Quantum State Diffusion

QUANTUM STATE DIFFUSION

Ian Percival

CAMBRIDGE UNIVERSITY PRESS

This is the ®rst book devoted to quantum state diffusion (QSD) and its applications to open quantum systems and to the foundations of quantum mechanics. It is suitable for readers with basic understanding of quantum theory. Recent experiments with detailed control over individual quantum systems have changed the face of quantum physics. These systems include atoms at the low temperatures attained by the 1997 Nobel Laureates, they include entangled photons in cavities, and they include the quantum systems used in new and future technologies such as quantum cryptography and quantum computation. The experiments have led to a revival of interest in the foundations of quantum mechanics. QSD is used both as a theoretical and computational tool to study these systems, and as the basis of new approaches to the foundations. Starting with classical Brownian motion in one dimension, this book leads the reader by easy stages into QSD as the theory of a continuously changing (open) quantum system that interacts with its environment. It then uses the same theory to analyse some modern alternative approaches to the measurement of quantum systems. The book will interest graduate students and researchers in quantum mechanics and its applications, including quantum optics, quantum statistics, the philosophy of quantum theory and theoretical molecular biology.

This page intentionally left blank

QUANTUM STATE DIFFUSION

This page intentionally left blank

QUANTUM STATE DIFFUSION Ian Percival Queen Mary and West®eld College, London

PUBLISHED BY CAMBRIDGE UNIVERSITY PRESS (VIRTUAL PUBLISHING) FOR AND ON BEHALF OF THE PRESS SYNDICATE OF THE UNIVERSITY OF CAMBRIDGE The Pitt Building, Trumpington Street, Cambridge CB2 IRP 40 West 20th Street, New York, NY 10011-4211, USA 477 Williamstown Road, Port Melbourne, VIC 3207, Australia http://www.cambridge.org © Ian Percival 1998 This edition © Ian Percival 2003 First published in printed format 1998

A catalogue record for the original printed book is available from the British Library and from the Library of Congress Original ISBN 0 521 62007 4 hardback

ISBN 0 511 00903 8 virtual (netLibrary Edition)

To Jill

This page intentionally left blank

Contents Acknowledgements

xii

1. 1.1 1.2 1.3

Introduction Description Summary Literature on related ®elds

1 1 2 5

2. 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9

Brownian motion and Itoà calculus Brownian motion and QSD Probabilities Itoà calculus Using Itoà calculus Drift and the interval  Correlation and ensemble covariance Complex diffusion Time scales and Markov processes Many ¯uctuations

6 6 11 12 14 15 17 19 20 21

3. 3.1 3.2 3.3 3.4 3.5 3.6 3.7

Open quantum systems States of quantum systems Ensembles of quantum systems Entanglement Open systems Measurement and preparation The boundary problem Quantum expectation and quantum variance

24 24 27 31 35 36 40 41

4. 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8

Quantum state diffusion Master equations QSD equations from master equations Examples Projectors Linear unravelling Other ¯uctuations QSD, jumps and Newtonian dynamics The circuit analogy

44 44 47 51 61 62 63 64 65

ix

x

Contents

5. 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8

Localization Measurement and classical motion Quantum variance and covariance, ensemble localization Quantum measurement Dissipation Channels and statistical properties Localization theorems Proof of dispersion entropy theorem Discussion

67 67 70 72 74 76 77 81 82

6. 6.1 6.2 6.3 6.4 6.5 6.6 6.7

Numerical methods and examples Methods Localization and the moving basis Dissipative quantum chaos Second-harmonic generation Continuous Stern±Gerlach Noise in quantum computers How to write a QSD program

85 85 88 89 93 94 95 97

7. 7.1 7.2 7.3 7.4 7.5 7.6 7.7

Quantum foundations Introduction Matter waves are real Niels Bohr and Charles Darwin Quantum theory and physical reality Preparation of quantum states Too many alternative theories Gisin condition

103 103 105 106 107 109 112 115

8. 8.1 8.2 8.3 8.4 8.5 8.6 8.7

Primary state diffusion ± PSD First approach ± SchroÈdinger from diffusion Decoherence Feynman's lectures on gravitation Second approach ± spacetime PSD Geometry of the ¯uctuations Matter interferometers Conclusions

119 119 123 124 125 129 129 132

Contents

xi

9. Classical dynamics of quantum localization 9.1 Introduction 9.2 Classical systems and quantum densities 9.3 Quantum expectations and other properties of densities 9.4 Probability distributions and means 9.5 Elementary density diffusion 9.6 Generalization 9.7 Density entropy decreases 9.8 Localization for wide open systems 9.9 Localization of a particle in a medium 9.10 Discussion

134 134 136 139 141 142 144 147 149 150 154

10. 10.1 10.2 10.3 10.4 10.5 10.6 10.7

155 155 156 160 161 162 166 169

Semiclassical theory and linear dynamics Classical equations for open systems Semiclassical theory of ensembles Semiclassical theory of pure states Localization regime Liner phase space transformations and squeezed states Linear dynamics and the linear approximation Summary and discussion

References

170

Index

179

Acknowledgements This book could not have been written without generous help from many people and institutions. I had help from my collaborators and my hosts, Gernot Alber, John Briggs, Tod Brun, Carl Caves, Terry Clark, Predrag Cvitanovic, Lajos DioÂsi, Barry Garraway, Jonathan Halliwell, Peter Knight, William Power, Peter Richter, RuÈdger Schack, Tim Spiller, Walter Strunz and particularly Nicolas Gisin, whose in¯uence can be seen throughout. There were others who willingly provided information or comments on early drafts, including Bobby Acharya, John Charap, Dave Dunstan, Artur Ekert, Gian-Carlo Ghirardi, Lucien Hardy, Frederick KaÂrolyhaÂzy, Philip Pearle, Tullio Regge, Antonio Rimini, Steve Thomas, Graham Thompson, Dieter Zeh and many more. Bob Jones, Jagjit Dhaliwal and Nelson Vanegas provided indispensable help in overcoming problems with computers and Steve Adams prepared the ®gures. Relevant ®nancial support came from the Alexander von Humboldt Foundation, the UK Engineering and Physical Sciences Research Council and the Leverhulme Foundation. I was lucky to have Simon Capelin, Jo Clegg, Mairi Sutherland and Ian Sherratt of Cambridge University Press, and also Meg Dillon and Jill Percival, to help with the editing and see the book through production, and to have Simon's encouragement and advice. Thanks to you all. Ian Percival

1 Introduction

1.1 Description Recent experiments have changed the face of quantum physics. They include the detailed control over the states of individual quantum systems such as atoms and photons, entanglement experiments such as tests of Bell's inequalities and also technology based on entanglement, such as quantum cryptography. They include laser cooling to picokelvin temperatures and advances in matter interferometry, leading to measurements of unprecedented precision. These new experiments need new theories, in particular an analysis of those individual quantum systems which have a signi®cant interaction with their environments, that is individual open quantum systems. The experiments expose weaknesses in the usual interpretation of quantum mechanics, and have revived interest in alternative quantum theories. They suggest precision measurements to test alternative quantum theories experimentally and to probe dynamics on Planck scales, such as times of about 10ÿ43 s. This book is about the physics of open quantum systems and the foundations of quantum mechanics. In 1984 Nicolas Gisin made a theoretical connection between them [68, 67], which many authors have developed since then. Quantum state diffusion, or QSD, is a mathematical theory that applies to both. Questions concerning the theoretical and numerical solution of speci®c practical problems, like the cooling of atoms to very low temperatures, or the effects of noise on a quantum computer, or the motion of a large molecule like a protein in water, are commonly separated from questions on the foundations of quantum mechanics, like the nature of quantum measurement, or the reality of matter waves in more than three dimensions. In QSD they are closely connected. To many physicists this looks bizarre. 1

2

Introduction

Nevertheless QSD can be applied to both. Applications and foundations have helped one another. Numerical solutions of QSD equations for open systems have provided insight into alternative quantum theories by showing in detail how quantum states can localize during a measurement. And the localization property of QSD, which arose from the study of quantum measurement, is crucial to the ef®cient numerical solution of some problems of open quantum systems. Experimental tests of alternative quantum theories are dif®cult because of noise from the environment, which can be analysed using the theory of open quantum systems and the practical application of QSD. There is a related convergence of two trends in physics previously quite distinct: the quantum measurement problem considered by physicists concerned with quantum foundations, and the quantum measurement process as considered pragmatically by experimenters and their theoretical colleagues looking for intuitive pictures and methods of computation. The following chapters concentrate on two themes, with little divergence from them: the application of quantum state diffusion to open systems and to the foundations of quantum mechanics. Readers who are interested in related ®elds will ®nd a guide to the literature at the end of this introduction. Most physicists working in quantum optics or condensed matter or molecular science naturally do not want to be involved with the foundations. Neither do most physicists or philosophers of science who work on the foundations want to bother too much about numerical and computational methods. The book takes account of this. Also, some of the more mathematical derivations could easily be skipped by readers who are prepared to take the results on trust. The connections between foundations and applications are emphasized here in the introduction, but not in the remaining chapters. 1.2 Summary Chapters 2 and 3 provide an introduction to the elementary theories of stochastic processes and open quantum systems. The next two chapters on general QSD theory and localization are for all readers. Open systems and quantum measurement are used as examples. Chapter 6 is on QSD as a numerical method for solving open system problems, and provides an introduction to a standard QSD computer program that is available on the World Wide Web. Chapters 7 and 8 are on the foundations of quantum theory. Chapters 9 and 10 contain more specialized theory topics, including classical and semiclassical QSD and the theory of open systems with linear dynamics. The theory in the book is illustrated by many examples.

1.2 Summary

3

The dependence of chapters on one another suggests reading sequences given by 1 ! 2 ! 3 ! 4 !5 ! 6 5!7!8 4 !9 ! 10; but much of the introductory and less mathematical parts can be read independently, particularly in chapters 7 and 8 on quantum foundations. Chapter 2 introduces open systems, the Itoà calculus and the Itoà form of Langevin stochastic differential equations from scratch, using Brownian motion as an example. Brownian motion is also used to illustrate the importance of time scales in Markov processes, and to introduce variance and covariance for ensembles. It assumes a nodding acquaintance with elementary probability theory. Chapter 3 brings in the relevant quantum theory and notation, with the emphasis on open systems, for which interaction with the environment is signi®cant and can be represented using ensembles with probabilities. It requires a background of elementary quantum mechanics in Dirac notation, including the picture of a quantum state as a vector in Hilbert space. Some properties of the density operator are illustrated for two-state systems by motion on and inside the Bloch sphere. The usual de®nition of a quantum measurement is generalized. Quantum variance and covariance for pure states of individual quantum systems appear. Chapter 4 uses the methods introduced in the previous two chapters to derive the Langevin-Itoà quantum state diffusion (QSD) equations for the quantum states of an individual open system from the master equation for the density operator of an ensemble. Alternative forms of the equations and the connection with quantum jumps are discussed. A gallery of graphs illustrates some solutions for simple examples. Chapter 5 is on localization. This is the property of QSD that makes the link between classical and quantum mechanics, which provides a dynamics for the process of quantum measurement. It also helps practical computations. Localization is described in terms of variances and entropies which decrease in the mean, and much of the chapter is devoted to showing this. Some readers may want to skip the derivations and just look at the results. Chapter 6 contains a description of numerical methods for the solution of QSD equations, especially the moving basis, which uses the localization property of QSD. It contains a user's guide to a program library, available

4

Introduction

on the World Wide Web. The methods are illustrated by some examples, including the Duf®ng oscillator, second harmonic generation, particles in traps, and quantum computers. The chapter assumes some knowledge of numerical methods and computer modelling, but not the details of the C++ language used for the standard computer library. The results of the computations may be of wider interest. Chapters 7 and 8 are on the foundations of quantum theory. They present some of the merits and problems of Bohr's Copenhagen interpretation and of those alternative quantum theories that are related to QSD. Some of the material in these chapters assumes an understanding of the physical implications of special and general relativity, but not the detailed theory. The part on experimental tests is written with experimenters in mind. Chapter 7 gives three reasons for the current interest in alternative theories, one theoretical and two experimental. The dif®culties of quantum theory are compared with those of early atomic physics and of the origin of biological species. Bell's conditions for a good quantum theory are stated, and it is shown that neither the Copenhagen interpretation, nor the environmental QSD of the earlier chapters satis®es them. There is a brief account of the development of some good quantum theories that are related to QSD, and a demonstration that some modi®cations of SchroÈdinger's equation are dif®cult to detect, so that there is a great variety of possible good theories, despite an important constraint. Chapter 8 shows how to restrict the possibilities by appeal to general principles and other ®elds of physics. Primary state diffusion is derived independently from two very different sets of principles. Firstly it is assumed that state diffusion is primary and the ordinary deterministic SchroÈdinger evolution is derived from it. This results in energy localization, whose merits and faults are discussed. The second derivation depends on assumed spacetime ¯uctuations on a Planck scale, due to quantum gravity. Special relativity is applied to the ¯uctuations, and is needed to produce position localization, but the dynamics of the system remains nonrelativistic. Atom interferometry experiments on spacetime ¯uctuations and the physics of the classical-quantum boundary are discussed, together with a brief account of the theoretical and experimental problems and prospects. Chapters 9 and 10 provide a phase space picture of the evolution of individual open quantum systems. Most of it needs a deeper understanding of classical dynamics in phase space than the rest of the book. Those who like to think classically will ®nd a helpful alternative picture of QSD in these chapters. They explain the classical dynamical theory of quantum localization, the

1.3 Literature on related fields

5

semiclassical limit of QSD, and the theory of localized systems with linear dynamics. 1.3 Literature on related ®elds This book is about the application of quantum state diffusion to the theory of open quantum systems and to the foundations of quantum theory. It deals either brie¯y or not at all with many important related ®elds, for which this section provides a brief guide. The treatment of quantum measurement in ordinary quantum theory textbooks does not deal adequately with continuous, repeated or incomplete measurements, such as continuous monitoring of quantum systems by photon counting in quantum optics, or nondestructive quantum measurements designed to detect gravitational waves. These continuous measurements can be handled adequately within the framework of ordinary quantum theory, as shown by Davies and Srinivas [26, 142, 141] and Mensky [94]. This provides a basis for the theory of quantum jumps [22, 121] and for the many practical applications of quantum trajectory methods using jumps, which were developed independently, with references given in section 4.7. Continuous measurement theory for gravitational waves is described in [19]. The mathematical theory of the stochastic evolution of quantum states has been developed by Hudson and Parthasarathy [85, 102], Belavkin [9], Barchielli [4] and their collaborators. John Bell's book [10] provides a stimulating introduction to the foundations of quantum theory. Wheeler and Zurek's collection of classic papers [154] is invaluable. The pilot wave theory of de Broglie and Bohm is treated in two recent books [18, 84]. In his book Bell points to similarities between this theory and the theory of Everett [49, 153]. Bohm discussed the role of causality and probability in quantum mechanics in [16], whereas [93] provides a non-mathematical introduction to the problems of quantum-nonlocality and relativity. The problems of quantum theory and gravity are mostly well above the level of this book, but the expositions in [108] are clearer than most.

2 Brownian motion and Itoà calculus 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9

Brownian motion and QSD Probabilities Itoà calculus Using Itoà calculus Drift and the interval  Correlation and ensemble covariance Complex diffusion Time scales and Markov processes Many ¯uctuations

A Brownian particle is an open classical system, and quantum state diffusion (QSD) is a theory of open quantum systems, so there is a parallel between the physics and mathematics of classical Brownian motion and the physics and mathematics of quantum state diffusion. But Brownian motion is much simpler, so it is used to introduce the properties of open systems, the Itoà calculus and the use of diffusion to probe small scales. All of these are important for QSD. 2.1 Brownian motion and QSD The interaction of an open system like a Brownian particle with its environment signi®cantly affects its dynamics. The motion is stochastic, so the future state of the system is not uniquely determined by its present state. The probability of a given future state is uniquely determined. General open systems are de®ned in section 3.4. Brownian motion is stochastic motion of an individual open classical system, in contrast to the determinism of Hamiltonian dynamics. Quantum state diffusion is stochastic motion of an individual open quantum system, in contrast to the determinism of SchroÈdinger dynamics. In QSD the quantum states diffuse in Hilbert space like Brownian particles diffusing in ordinary position space. For example, the system may be an atom in a radiation ®eld, a radiation ®eld in the presence of atoms, a molecule in a liquid, or a quantum computer in the presence of noise. It may be a quantum system which is being measured. In QSD this diffusion is the source of quantum ¯uctuations and the indeterminism of quantum physics. 6

2.1 Brownian motion and QSD

7

Brown did not discover Brownian motion. Anyone looking at water through a microscope is likely to see little things wiggling around, and according to Brown they were seen by Leeuwenhoek, who lived from 1632 to 1723. In 1828 Brown established Brownian motion as an important phenomenon, demonstrated that it was present in inorganic matter and refuted by experiment various spurious mechanical explanations. In the late nineteenth century there were good scientists like Ostwald and Mach who did not believe in the reality of atoms. For them, atomic theory was just a convenient picture or model which could be used in physics and chemistry to explain the properties of observable macro systems, and those who believed in the reality of atoms were misguided. Einstein's 1905±1908 theory and the later experiments of Perrin established the reality of atoms without question. Brownian motion is a diffusion process, and because of this, macro scale experiments on Brownian motion could be used to probe the atomic scale, even though the details of atomic dynamics were not known at the time. These experiments are much easier than the more direct methods like X-ray diffraction, scattering experiments and ultramicroscopes. Brown's systematic studies were completed long before such methods were possible. QSD is also a diffusion process. It is suggested in chapter 8 that quantum state diffusion might be used to determine quantities on the Planck scales of length and time, which are smaller than atomic scales by more than twenty orders of magnitude. The role of QSD in quantum theory today parallels the role of Brownian motion in the atomic theory of a century ago. The books of Einstein and Nelson give good accounts of the early work on Brownian motion [46, 99]. There is a parallel between the theory of Brownian motion and the theory of quantum state diffusion. A Brownian particle moves along a crooked path r…t† in position space. This crooked path is a solution of a stochastic differential equation, the Langevin equation. Langevin introduced differential equations with stochastic forces, which will be treated more consistently in section 2.3 using ItoÃ's stochastic calculus. The state of an ensemble of Brownian particles is represented by a density …r; t† that is the solution of a linear deterministic partial differential equation, called the Fokker-Planck equation. According to QSD, a single open quantum system moves along a crooked path or trajectory j …t†i in the state space, that is a solution of a stochastic differential equation, the QSD equation. The evolution of an ensemble of quantum systems is represented by a density operator q…t† that is a solution of

8

Brownian motion and Itoà calculus

a linear deterministic equation, called the master equation, which may be a matrix differential equation or a partial differential equation. There is also a parallel in the computational methods that are used today. For complicated Brownian processes, such as the motion of charged Brownian particles in an electric ®eld that varies in space and in time, it is often easier to solve the Langevin equation numerically for a sample of particles than to solve the Fokker-Planck equation for the density. The ®rst method is a Monte Carlo method. The Monte Carlo method is less accurate, but it can be used where the Fokker-Planck equation is insoluble in practice. Similarly, for open quantum systems that require many basis states, it is often possible in practice to solve the QSD equation numerically for a sample, when it is impossible to solve the master equation for the density operator, as shown in chapter 6. These parallels are well established. There are others which are less well established. For example, the theory of Brownian motion clari®ed the foundations of kinetic theory, whereas QSD clari®es the foundations of quantum mechanics. Theory and experiments on Brownian motion provided access to the physics of atomic scales, whereas the primary state diffusion theory and matter interferometer experiments of chapter 8 might provide access to the physics of the Planck scales. So Brownian motion and its theory provide a good introduction and background to the physics and mathematics of QSD, despite the important differences between them. For simplicity, consider only the vertical motion of a Brownian particle. First suppose that the density of the particle is the same as the ¯uid around it. Then there is no drift due to gravity. The displacement of the particle ¯uctuates due to collisions of the Brownian particle with molecules. This process takes place on the atomic scale, producing random upward and downward displacements, resulting in a diffusion away from its initial height. The mean vertical displacement x…t† of the particle from its initial height is zero, M x…t† ˆ 0;

…2:1†

where M will always be used to represent the mean over stochastic ¯uctuations. `Expectation' will not be used for such ¯uctuations, as it is needed for quantum expectations. As in statistical mechanics, a mean is a property of a large ensemble of individual systems. For our example, the mean vertical

2.1 Brownian motion and QSD

9

displacement is a property of an ensemble of Brownian particles that start from the same height. The mean square displacement from the initial position is proportional to the time, M x2 …t† ˆ a2 t;

…2:2†

where by convention a is a positive constant. The value of a depends on the mass of the particle, the viscosity of its ¯uid environment and the temperature, but this does not concern us here. A typical displacement due to the diffusion is proportional to the square root of the time, and this can be expressed in terms of the root mean square (RMS) deviation,
x…t† ˆ

q p M x2 …t† ˆ a t:

…2:3†

When the density of the Brownian particle is greater than the density of the ¯uid through which it moves, there is an additional drift due to gravity. Its vertical displacement is then made up of two parts, drift and stochastic ¯uctuations, vertical displacement = drift + stochastic fluctuations.

…2:4†

Here and throughout the book a ¯uctuation will always be a stochastic variable with zero mean. It is only then that there is a unique decomposition of the motion into separate drift and ¯uctuation components. Here a ¯uctuation is a displacement and not a velocity. The drift, due to the force of the Earth's gravitational ®eld, is uniformly downwards. In a short time interval t, the mean square displacement M …x†2 due to the diffusion is proportional to t, so a typical displacement is proportional to the square root of the time interval. The drift is proportional to the time interval t itself. The absence of the expected acceleration term is explained in section 2.8. Figure 2.1 illustrates simulated Brownian motion with drift. Figure 2.1(a) gives the height. Over very short periods of time it is dif®cult to see the drift because it is masked by the ¯uctuations, whereas over longer periods the drift becomes relatively more important. Over very long periods it may be dif®cult or impossible to distinguish thep¯uctuations at all. Figure 2.1(b) shows the  root mean square displacement M …x2 † from the initial height at t ˆ 0 for a sample of 100 Brownian particles. This is not a deviation from the mean. For

10

Brownian motion and Itoà calculus

Figure 2.1a The height of a Brownian particle as a function of time.

Figure 2.1b The RMS deviation from the initial height for a sample of 100 particles.  is de®ned in the text.

2.2 Probabilities

11

small times it looks roughly parabolic, typical of the square root dependence of the ¯uctuations without the drift. For longer times it looks linear, typical of drift without ¯uctuations. The ®gure also shows the time interval  that marks the approximate boundary between the smaller time intervals for which the diffusion dominates and the larger time intervals for which the drift dominates. The measured value of  provides valuable information about atomic scales. For times T much longer than , the ¯uctuations are typically about …=T†1=2 of the displacement due to the drift. This is the ¯uctuation factor, which is de®ned for all processes in which there are both ¯uctuations and drift. It is important for the experimental tests of spacetime ¯uctuations proposed in chapter 8.

2.2 Probabilities A Brownian particle is a special case of a system interacting with its environment. The state of the environment cannot be represented exactly, but, as usual in statistical mechanics, the probability of a given state can be represented. Consequently, interaction between system and environment produces indeterminate stochastic evolution of the system, represented by an ensemble with a probability distribution in space for the Brownian trajectories. The notation used here for probabilities is easily extended to quantum systems. Let Pr…x† be a probability distribution for a discrete or continuous variable x. Then the total probability for any value of x must be 1, so X

Z Pr…x† ˆ 1

or

dx
Pr…x† ˆ 1;

…2:5†

x

where the sum or integral is taken over all possible values of x. Such sums and integrals are so common that it is convenient to use a general notation for both which is consistent with the corresponding quantum notation. The trace Tr is the sum or integral over all possible values of a variable x and so Tr Pr…x† ˆ 1:

…2:6†

The variable x can represent a dynamical variable of some classical system, like the height of a Brownian particle. If f …x† is some property of the system,

12

Brownian motion and Itoà calculus

like the vertical distance of the particle from some point, then the mean of f for the probability distribution Pr…x† is M f ˆ Tr Pr…x† f …x†:

…2:7†

The ensemble variance or mean square deviation of f is 2 … f † ˆ M … f ÿ M f †2 ˆ M … f 2 † ÿ …M f †2 :

…2:8†

This is always positive or zero and its square root
f is the standard deviation. All the variances of this chapter are classical ensemble variances. In chapter 3 we will meet the corresponding quantum variances. The probability distribution for y ˆ f …x† is Pr…y† ˆ Trx Pr…x† … f …x† ÿ y†;

…2:9†

where the  function is the Dirac or Kronecker , depending on whether x is continuous or discrete.

2.3 Itoà calculus Itoà and Stratonovich have done for stochastic dynamics what Newton and Leibniz did for ordinary dynamics, by treating the small displacements as differentials. This results in a very elegant theory and differential calculus for continuous stochastic processes that satisfy Langevin equations, sometimes known as Wiener processes. This and other stochastic methods are described for physicists in the books of Gardiner [55, 56]. Remarkably, Itoà calculus is a differential calculus of nondifferentiable functions. When a Brownian particle with vertical displacement x diffuses with no drift, the mean of the vertical distance traversed is zero and the mean square or variance is proportional to the time. So for small displacements dx, M dx…t† ˆ 0;

M ‰dx…t†Š2 ˆ a2 dt;

…2:10†

where M represents a mean over all possible displacements dx…t† and a is a positive constant. The small displacements are treated as differentials by ItoÃ, and their higher powers are neglected, as in the ordinary Leibniz differential calculus, but the rules are different. For more complicated stochastic pro-

2.3 Itoà calculus

13

cesses, it is convenient to introduce a standard normalized stochastic differential dw, with zero mean and variance equal to the time, so that M dw ˆ 0;

M …dw†2 ˆ dt:

…2:11†

For these differential ¯uctuations, only these means are signi®cant. It does not matter whether the ensemble of stochastic differentials is made up of steps of ®xed length with equal probability of going to the left or to the right, or of steps with a Gaussian distribution, or a rectangular distribution. They all have the same effect, provided that they satisfy (2.11). It is implicitly assumed that ¯uctuations at different times are independent of one another. This is the Markov assumption, discussed in more detail in section 2.8. Brownian motion with no drift is then given by the solution of the simplest Langevin-Itoà differential equation dx…t† ˆ a dw…t†:

…2:12†

In this equation, and in general for Itoà stochastic differential equations, it is assumed that, at time t, the trajectory x…s† with 0  s  t is already determined, so that only the new ¯uctuation dw…t† need be considered. All the particles of the ensemble are at the initial point x…t† at time t, but they are usually at different points at time t ‡ dt. So for a single step of the equation the mean M refers to the ensemble of the current ¯uctuations dw…t† only. Functions of t and x…t† have a ®xed value, and so may be taken outside the mean. The properties of more general ensembles, formed from all the ¯uctuations over ®nite time intervals, can be derived from the properties of these elementary ensembles, in which the initial ensemble consists of systems in the same state. That is why most of the Itoà equations in this book have no mean M on the right hand side. There is no need for it when all the initial states are identical, and general equations can be obtained from the elementary ensembles. In ordinary differential calculus the whole theory can be based on ®rstorder differentials, higher orders being neglected by comparison with the ®rst order. The Itoà differential calculus is also based on lower-order differentials, but they are not just of ®rst order. Stochastic differentials are of order …dt†1=2 , so the squares of stochastic differentials like dw cannot be neglected. Squares of ordinary differentials like dt can be neglected as usual, and so can products of the form dt dw, which are of order …dt†3=2 . The contributions of drift and diffusion terms can be considered independently, because cross

14

Brownian motion and Itoà calculus

terms are negligible, which simpli®es the analysis of the diffusion equations considerably. A stochastic differential of order dt whose mean is zero has variance of order …dt†2 and therefore this is also negligible. In the Itoà calculus all such
negligible differentials are set to zero. The variance of ÿ M …dw†2 ÿ …dw†2 is of order …dt†2 , which can be neglected, so if we want to save notation, we can write …dw†2 ˆ dt;

…2:13†

without the mean that appears in equation (2.11). Following the rules of the Itoà calculus, the differential of a function f of a stochastic variable x and a product of two functions f and g are df ˆ dx
f 0 ‡ 12…dx†2
f 00 ;

d… fg† ˆ f dg ‡ df
g ‡ df
dg:

…2:14†

The ordering of the terms has been retained on the right side of the second equation, because QSD sometimes has noncommuting ¯uctuations. 2.4 Using Itoà calculus The increase with time of the variance of the height of a Brownian particle without drift is a simple example of the Itoà calculus. Suppose for simplicity that at t ˆ 0 the displacement is x ˆ 0. Then the variance at some later time is 2 …x† ˆ M x2 . The change in the variance after time dt is d2 …x† ˆ d M x2 ˆ M d…x2 †

(change of the mean is mean of the change).

(2.15)

Using equation (2.14) for the differential of a product, d2 …x† ˆ M 2xdx ‡ M …dx†2 :

…2:16†

The initial displacement x is independent of the following ¯uctuation, so x can be taken outside the mean, and M xdx ˆ x M dx ˆ xa M dw ˆ 0; d2 …x† ˆ M …adw†2 ˆ a2 dt;

 …2:17†

2.5 Drift and the interval s

15

where the important term is the square of a stochastic differential, which cannot be neglected in the Itoà calculus. The averaged picture of Fokker and Planck does not represent the motion of the individual particles directly, but uses the probability distribution …x; t† of an ensemble of particles. If there is no drift, its differential equation follows from the stochastic differential equation (2.12), using the Itoà calculus. For a particular ¯uctuation dw, the distribution at time t ‡ dt is given by shifting …x; t† by dx, giving …x; t ‡ dt† ˆ …x ÿ dx; t† ˆ …x ÿ adw; t† ˆ …x; t† ÿ a0 …x; t†dw ‡ 12 a2 00 …x; t†…dw†2 ;

…2:18†

where the prime means differentiation with respect to x. For the whole ensemble we take the mean over all possible ¯uctuations. Now M dw ˆ 0, M …dw†2 ˆ dt and …x; t ‡ dt† ˆ M …x ‡ dx; t† ˆ …x; t† ‡ 12 a2 00 …x; t†dt; so

2 00

d ˆ 12 a  …x; t†dt

…2:19†

(Fokker-Planck):

The solution of (2.19) is the set of Gaussians illustrated in ®gure 2.2. These two derivations are typical of many that appear in QSD. Because the ordinary differential calculus is used so much, most people ®nd it dif®cult to avoid neglecting some of the squared differentials at ®rst. This problem is overcome in the alternative but equivalent Stratonovich calculus [54], at the price of a slightly more dif®cult interpretation of the results, but the Itoà form is used here. 2.5 Drift and the interval s When the density of the particles is greater than their ¯uid environment, the Earth's gravitational ®eld adds a drift velocity v to the vertical diffusion, giving the stochastic differential equation dx ˆ v dt ‡ a dw;

…2:20†

where the coef®cient of dt gives the drift velocity and the coef®cient of dw the magnitude of the diffusion. If x is replaced by x ÿ M x, equation (2.20)

16

Brownian motion and Itoà calculus

Figure 2.2 Gaussian solutions of the Fokker-Planck equation for Brownian motion without drift.

reduces to equation (2.12), so the variance of the displacement x is unaffected by the drift. Also for more general differential stochastic processes, in which x represents the state of some dynamical system, the drift is the term containing dt and the diffusion terms contain ¯uctuations like dw. For suf®ciently small times, the diffusion, which is proportional to …dt†1=2 , dominates the drift, which is proportional to dt. The approximate boundary between the smaller time intervals dominated by the diffusion and the longer time intervals dominated by drift is obtained by supposing that equation (2.20) remains valid for ®nite time intervals, and equating the increments due to ¯uctuation and drift, giving v ˆ a 1=2 ;

 ˆ …a=v†2

(boundary interval).

This is the boundary interval illustrated in ®gure 2.1 on page 10.

…2:21†

2.6 Correlation and ensemble covariance

17

2.6 Correlation and ensemble covariance This section follows from section 2.2 on probabilities. In any ensemble of systems, two dynamical variables f and g may or may not be correlated. If they are uncorrelated then their joint distribution is the product Pr… f ; g† ˆ Pr… f †Pr…g†

…2:22†

and a knowledge of the value of f then gives no information about the value of g. The ensemble covariance of f and g, … f ; g† ˆ M … f ÿ M f † …g ÿ M g† ˆ M … fg† ÿ …M f † …M g†

…2:23†

is de®ned in terms of the deviation of f and g from their means. It is a bilinear measure of the correlation between f and g. It is zero if f and g are uncorrelated. If it is nonzero then they are correlated. But f and g can be correlated nonlinearly, in which case the covariance may be zero, despite the correlation. An example of this is when the variables x and y are distributed uniformly on a circle in the …x; y†-plane, represented by x ˆ cos ;

y ˆ sin ;

…2:24†

where  is distributed uniformly between ÿ and . The variables x and y are then strongly correlated, but their covariance is zero. However, if the covariance of every function of f and every function of g is zero, then f and g are uncorrelated. The covariance of f with itself is just the variance of f . … f ; f † ˆ 2 … f †:

…2:25†

We often need to deal with the statistical properties of complex variables f ; g. There is a natural de®nition of the variance, which is … f ; f † ˆ 2 … f † ˆ M … f  f † ÿ …M f † …M f † ˆ M j f j2 ÿ jM f j2 ˆ 2 … fR † ‡ 2 … fI †;

…2:26†

where zR and zI are the real and imaginary components of a complex quantity z. For general probability distributions in complex f and g, the properties of the covariance are not so simple, but we will not be concerned with general

18

Brownian motion and Itoà calculus

distributions, only with those distributions where multiplication of the important variables by a phase factor u leaves all probabilities, including joint probabilities, unchanged, so that Pr…uf ; ug† ˆ Pr… f ; g†

…juj2 ˆ 1†:

…2:27†

In that case there is a natural generalization of the de®nition of ensemble variance to ensemble covariance, given by … f ; g† ˆ M … f  g† ÿ …M f † …M g† ˆ … fR ; gR † ‡ … fI ; gI †;

…2:28†

which has the same basic properties as the ensemble covariance of real variables. Correlation between systems is particularly important. If A and B are two systems, and AB is the combined system, then A and B are uncorrelated in an ensemble if the state x of A is uncorrelated with the state y of B, so PrAB …x; y† ˆ PrA …x†PrB …y†:

…2:29†

It follows that if the covariance of every function of the state of A with every function of the state of B is zero, the systems are uncorrelated, and conversely. Classical probabilities, correlations, variances and covariances are properties of classical ensembles, not of individual systems, even when the ensemble is not mentioned explicitly. We will see in later chapters that this is not true of quantum variances and covariances. Two systems that have not interacted either directly or indirectly are always uncorrelated. When they do interact, they almost always become correlated. We often want to study the behaviour of one of the systems on its own, even when it is continually interacting with another system. For example, an open system is continually interacting with its environment, and the environment is usually so big and complicated that it cannot be studied in detail. Brownian motion is an example. In such cases we are only interested in the reduced probability distribution for the system itself. Let A be the system and B the environment. Write TrB for the integral over all the states y of the environment. Then the reduced distribution for A is PrA …x† ˆ TrB PrAB …x; y†:

…2:30†

If we know something about the distribution of B, for example that it is in thermal equilibrium at a ®xed temperature, and we know the dynamics of the

2.7 Complex diffusion

19

interaction between A and B, then we can often derive stochastic equations of motion for A, and deterministic evolution equations for the probability distribution PrA …x†. The dynamics of quantum mechanical amplitudes for pure quantum states has some features in common with the dynamics of classical probabilities for ensembles of classical systems. Similarly quantum entanglement, described in section 3.3, resembles classical correlation in some respects. Despite these common features, quantum amplitudes and entanglement have some special properties of their own.

2.7 Complex diffusion For applications to quantum mechanics, isotropic diffusion in two real dimensions is best considered as one-dimensional complex diffusion in the complex z-plane illustrated in ®gure 2.3. The standard normalized complex stochastic differential is d ˆ dR ‡ idI ; where the real and imaginary parts satisfy

Figure 2.3 Isotropic diffusion in the complex z-plane.

…2:31†

20

Brownian motion and Itoà calculus

M dI ˆ M dR ˆ 0 M dI dR ˆ 0 2 ˆ M dI2 ˆ dt=2 M dR

(zero mean); (zero covariance); (normalized):

The equivalent complex form of the conditions is more convenient: M d ˆ 0 …d†2 ˆ …d ; d† ˆ 0 2

2

jdj ˆ  …d† ˆ dt

(zero mean); (zero covariance);

…2:33†

(normalized);

where because of the isotropy the statistics of the motion is unaffected if d is multiplied by a phase factor of unit magnitude. The equations (2.14) for the differential of a product and of a function apply to complex ¯uctuations as they do to real ones, but there is an important simpli®cation for complex ¯uctuations, because the square of a complex differential ¯uctuation is zero, so the second term in the expression for df vanishes. This leads to a simpler quantum state diffusion theory when complex ¯uctuations are used. The Langevin-Itoà equation of a particle with drift and diffusion at position z…t† in the complex plane is dz ˆ vdt ‡ ad:

…2:34†

Both v and a can be complex, but the physics is unaffected by the (gauge) transformation for which a is multiplied by a phase factor. This completes the classical theory needed for an understanding of elementary quantum state diffusion for open systems. Quantum state diffusion is a diffusion of quantum states represented by normalized state vectors in a complex Hilbert space, which is not so simple. But many of the methods and notation introduced in this chapter are used. 2.8 Time scales and Markov processes The Itoà equations for Brownian motion are not valid on all time scales. For example, if the environment of the Brownian particle is a gas, a molecule of the gas takes a ®nite time to collide with the particle. For classical collisions we could take the time interval Tcol to be the minimum time for the collision to produce half the total momentum transfer. During such time intervals the

2.9 Many fluctuations

21

motion looks smooth and cannot be represented by an Itoà equation like (2.12) or (2.20). There are correlations between the ¯uctuations over time intervals comparable to Tcol and smaller, so that on this scale the Markov assumption that there are no correlations is wrong. But for time scales that are long by comparison, the correlations are negligible and the Itoà equations can be used. Typically Tcol is smaller than the boundary time  by many orders of magnitude, and cannot be detected experimentally. In the theory of Brownian motion, a uniform drift has been assumed, which seems incompatible with Newton's second law of motion. A heavy body in air, like a skydiver's, ®rst accelerates. The resistance of the air then reduces her acceleration towards zero, so that she approaches a terminal drift velocity. Her Brownian ¯uctuations are negligible. For Brownian particles the period of acceleration is normally so short that it can be neglected. Only the terminal drift velocity and the ¯uctuations can be seen. For dense gases and liquids the time between collisions is very short compared with the collision time. However, for diffusion in gases of very low density, it may be very much longer, and when the mean free path is macroscopic, the Markov assumption and Itoà equations may not be valid on any scale. The quantum mechanics of the collisions complicates things further. It was Einstein's achievement to recognize that the relevant theory was usually independent of such complications. That is how he produced his theory more than two decades before any proper understanding of the elementary collisions which cause the ¯uctuations. Similar considerations apply to quantum state diffusion. The Markovian Itoà theory of quantum state diffusion is valid on time scales that are long compared with the basic time scale for interaction with the environment, but sometimes the correlations between successive ¯uctuations cannot be ignored, the evolution of the quantum system is non-Markovian, and QSD cannot be used. 2.9 Many ¯uctuations This section is a little bit more dif®cult than the rest of the chapter. It includes the classical theory needed for all but the simplest applications of QSD, and can be skipped at ®rst reading. De®nitions for single ¯uctuations are generalized to many ¯uctuations, which introduces some mathematical complications. Suppose there are K real ¯uctuations dw0k with zero mean. We cannot always rule out the possibility that the ¯uctuations dw0k are correlated, or even linearly dependent. Since powers of dw0k higher than squares are negli-

22

Brownian motion and Itoà calculus

gible, the means and variances and covariances provide all the statistical information about them, so M dw0k ˆ 0

(zero mean),

0 0 dt (general), …dw0k ; dw0k0 † ˆ M dw0k
dw0k0 ˆ Xkk

 …2:35†

0 0. thus de®ning Xkk The variance of a ¯uctuation is treated as a norm, and the covariance of two ¯uctuations as a scalar product. By taking linear combinations of the ¯uctuations we can transform them to a standard form, in which they are all normalized to dt and orthogonal. This makes an orthonormal set of ¯uctuations. Normalization is no problem, because ¯uctuations are always used with coef®cients, and unwanted constant factors can be absorbed into the coef®cients. Orthogonality is the problem. 0 0 can be diagonalized by an orthogonal transforThe symmetric matrix Xkk mation:

Xjj 0 ˆ

X kk0

0 tr 0 Ok0 j 0 ; Ojk Xkk

…2:36†

giving a new set of ¯uctuations that are linear combinations of the old ones: dw0j ˆ

X k

Ojk dw0k :

…2:37†

For the new ¯uctuations, the off-diagonal covariances are zero and, if some of the old ¯uctuations dw0k are linearly dependent, some of the new variances dw0j are zero too. The latter may be ignored, as they do not contribute to the diffusion. Suppose there are J remaining. Generally they are not normalized, but they can be normalized so that each variance is dt, giving a set of orthonormal ¯uctuations: M dwj ˆ 0 …dwj ; dwj 0 † ˆ M dwj
dwj 0 ˆ jj 0 dt

(zero mean); (orthonormality).

…2:38†

This is the standard normalization, in which the norm of the vector ¯uctuation fdwj g is Jdt. There is a direct generalization to diffusion in a complex vector space, which is needed for QSD. A general diffusion has correlated ¯uctuations dk0 and nonzero off-diagonal covariances, using the de®nition of complex

2.9 Many fluctuations

23

0 0 can covariance of section 2.6 in equation (2.28). The Hermitian matrix Xkk be diagonalized by a unitary change of basis. For the new ¯uctuations dk0 , the off-diagonal covariances are zero and, if some of the ¯uctuations dk0 are linearly dependent, some of the diagonal variances of dj0 are zero too, and there are only J < K independent ¯uctuations. These can be normalized. The new complex ¯uctuations have the standard form

M dj ˆ 0

…dj ; dj 0 † ˆ

M dj

(zero mean),


dj 0 ˆ jj 0 dt

(orthonormality).

…2:39†

The standard forms are not unique. Orthogonal transformations of fdwj g and unitary transformations of fdj g leave their properties are unchanged. These may be considered as different ¯uctuations or as a different representation of the same vector ¯uctuation in the linear ¯uctuation space. A general Itoà equation for a real variable x with many independent real ¯uctuations is dx ˆ vdt ‡

X j

aj dwj :

…2:40†

Let f …x; t† be a differentiable function of x and of t. Because it is differentiable with respect to t it does not depend explicitly on the ¯uctuations. The change in f …x; t† in a time dt is df …x; t† ˆ …@f =@x†dx ˆ …@f =@x†vdt ‡

X j

…@f =@x†aj dwj :

…2:41†

Consequently the contributions to df from the drift term and each of the ¯uctuation terms proportional to dwj are additive. This is the additivity rule, which is also true for complex x with complex ¯uctuations, when x is a vector, or even when it is a quantum state vector, as for QSD. The additivity rule is used extensively in QSD to obtain the theory of complicated systems from the theory of elementary systems.

3 Open quantum systems 3.1 3.2 3.3 3.4 3.5 3.6 3.7

States of quantum systems Ensembles of quantum systems Entanglement Open systems Measurement and preparation The boundary problem Quantum expectation and quantum variance

An individual open quantum system can be represented by a pure state, and the statistical properties of an ensemble of open systems by a probability distribution over pure states. The density operator which can be obtained from this distribution is more compact, but does not represent the dynamics of the individual systems of the ensemble. A density operator is also used to represent an open quantum system that is entangled with its environment as the result of interaction. Measurement and preparation are given general de®nitions in terms of the interaction of a quantum system with its environment. The boundary between system and environment is ambiguous. For a pure state there are quantum expectations, quantum variances and quantum covariances. 3.1 States of quantum systems A pure state of a quantum system with a state vector j i that satis®es the SchroÈdinger equation is a very useful ®ction. Even if the system is isolated, every system we know has interacted with other systems in the past, and so has become entangled with them, as described in section 3.3. Strictly speaking, no system is independent, and so no system can be properly represented by a state vector j i. All systems are open, with the exception of the whole universe. However, quantum systems are commonly represented by state vectors, theories based on this representation are very successful, and we use this representation too. The representation of the same pure state by a projection operator or projector P ˆ P ˆ j ih j; 24

…3:1†

3.1 States of quantum systems

25

is equivalent to state vector representation, except for a physically irrelevant external phase factor. We use both. The SchroÈdinger time evolution of a pure quantum state j i with Hamiltonian H is given by the differential equations j _ …t†i ˆ ÿ…i=hÿ†Hj …t†i

and

P_ ˆ ÿ…i=hÿ†‰H; PŠ;

…3:2†

from which it follows that the evolution over ®nite times is given by j …t†i ˆ exp…ÿiHt=hÿ†j …0†i

and

P…t† ˆ exp…ÿiHt=hÿ†P…0† exp…iHt=hÿ†: …3:3†

This evolution is linear. It is also deterministic, meaning that the present state determines future states uniquely. The norm h j i of the state vector and the corresponding trace Tr P of the projector are preserved. Linear norm-preserving transformations are unitary. Quantum state diffusion preserves the norm and the trace, but it is neither linear nor deterministic. It is nonlinear and stochastic. An open quantum system continually interacts with its environment. An ensemble of these systems is sometimes called a `mixed state' of the system and is represented by a density operator which is de®ned in the next section, 3.2. But in QSD we usually use the representation of a single state of the ensemble by a pure state j i. For open quantum systems, just like isolated ones, this is a very useful ®ction. We also use the density operator, but it plays a secondary role. The properties of systems with many orthogonal states are easier to understand when there are only two, as in the example of a spin one-half system. This chapter uses these `two-state' systems to illustrate mathematically and pictorially the properties of the more general systems that are needed for QSD. By convention, the states j‡i and jÿi with `up' and `down' spins in the ‡z and ÿz direction are used as the base vectors …1; 0† and …0; 1† respectively. These are the eigenvectors of the Pauli matrix rz . The eigenvectors of rx represent the spins in the x direction and of ry represent the spins in the y directions. The matrix representations of the operators and eigenvectors can be found in most quantum mechanics texts. The states and ensembles of `two-state' systems are best visualized in the three-dimensional linear space spanned by the projectors, which is not the same as the Hilbert space of the states themselves. The projectors are points

26

Open quantum systems

on the surface of the Bloch or Poincare sphere, illustrated in ®gure 3.1. The projectors P can be represented by 2  2 matrices. In the space of the projectors, the centre of the sphere, which is the origin of the coordinates, is at half the unit matrix, 12I. The vector representing the projector P is r~ ˆ …x; y; z† ˆ TrP~r;

where

~ ˆ …rx ; ry ; rz † r

…3:4†

and the projector corresponding to the vector r~ is ~ †: P ˆ 12…I ‡ r~
r

…3:5†

The projectors of the two eigenstates of a general spin-half matrix r are at 1 2…I  r†, which are on opposite sides of the sphere. A point on the surface of the sphere indicates the direction of the spin in three-dimensional position space. The up and down spins are at the north and south poles. Besides spin-half systems, other examples of two-state systems are polarization states of photons, the two states of an atom that resonate strongly with laser light of a single ®xed frequency, and the two occupation states for a fermion. For photons, the up and down spins correspond to vertical and

Figure 3.1 Projectors as points on the surface of the Bloch sphere.

3.2 Ensembles of quantum systems

27

horizontal polarizations of the photon, the eigenstates of rx to linear polarization at 458 and the eigenstates of ry to circular polarizations. For two-state atoms, the up and down spins correspond to the excited and ground states of the atom and the other spins to coherent combinations of the two atomic states. Two-state quantum systems, like two-state classical systems, can be used to represent binary numbers. The corresponding bits are called qubits, spoken as `kewbits'. Quantum systems of many qubits behave very differently from classical systems with many classical bits. For example, simultaneous parallel operations can be performed on qubits that are impossible with classical bits. This is the basis of modern quantum technology, including quantum communication, quantum cryptography and possible future quantum computers. QSD has been applied to quantum computers, which are discussed further in chapter 6. All two-state Hamiltonians have the form hÿ!r, where ! is an angular frequency and r is a spin operator in an arbitrary direction. The SchroÈdinger evolution of a pure quantum state is then given by uniform rotation on the surface of the Bloch sphere with angular velocity ! about an axis through the eigenstates of r. For H ˆ !rz , a state moves uniformly around a line of latitude. The book by Allen and Eberly [2] has a useful introduction to the Bloch sphere. 3.2 Ensembles of quantum systems Classical dynamics and SchroÈdinger dynamics are both deterministic, but there is an unavoidable stochastic behaviour where they interact. Because of this, probability is more important for quantum theory than it is for classical theories. A single system whose state is not perfectly known, or whose evolution is stochastic, is represented by an ensemble of systems with a probability distribution over pure states. One example is an open quantum system, like a molecule in a ¯uid. Another is a collection of many quantum systems, like many runs of an experiment. Depending on the type of ensemble, the probabilities can be represented by a continuous probability distribution Pr… † in the very large space of all quantum states j i with projectors P , or a discrete distribution, which consists of a set of probabilities Pr…j† for states j ji with projectors Pj . The states need not be orthogonal in either case. For simplicity, the theory is developed here for discrete distributions, which can be generalized without

28

Open quantum systems

dif®culty to continuous distributions, replacing sums by integrals where appropriate. Continuous distributions are discussed in [6]. For quantum state diffusion theory, the probability distribution is the most important representation of an ensemble of quantum systems. But it is not a common representation, because the measurable properties are all given by a simpler representation using density operators. The density operator and the corresponding vector for an ensemble of quantum systems are qˆ

X j

Pr…j†Pj ;

r~ ˆ Trq~r

…3:6†

and the point r~ is in the interior of the Bloch sphere unless the members of the ensemble are identical. Remarkably, ensembles with the same density operator cannot be distinguished from one another by measurements, however they may be performed, so that for many purposes such ensembles are equivalent. This equivalence is a major theme of this book. However, a density operator does not normally represent the individual systems of the ensemble. An exception is the very special type of ensemble which consists of identical pure states, for which the ensemble is represented by the pure state projector for that state. The probability of the system being found by a measurement in the state j i is TrqP ˆ h jqj i:

…3:7†

This property is sometimes used to de®ne the density operator. The probabilities must be positive or zero, and their sum for a complete set of orthonormal states is unity. In other words, the diagonal elements of any matrix representation of q must be non-negative and sum to unity. These two conditions are the positivity and trace conditions on the density operator. An ensemble of quantum systems may start as an ensemble of identical pure states and then evolve into an ensemble of differing pure states, just as an ensemble of Brownian particles starting at one place evolves into an ensemble at different places. Examples are quantum Brownian motion, radioactive decay, optical transitions in atoms and nuclei, and quantum measurements. All these processes can be treated using the theory of open quantum systems.

3.2 Ensembles of quantum systems

29

An ensemble of different pure states of a two-state system can be represented by a probability distribution on the surface of the Bloch sphere, or by a point in its interior representing the density operator. The point representing the density operator is a linear combination of the points on the surface, with weights given by their probabilities, like a centre of mass. If there are only two pure states in the ensemble, then the density operator lies on the line joining their projectors. The inverse process of unravelling a density operator into its component pure states is not unique, because the same density operator can be formed from many different combinations of pure states, as illustrated in ®gure 3.2, unless the ensemble consists of many copies of a single pure state. The SchroÈdinger dynamics of a probability distribution Pr… † over pure states is given directly by the evolution of the individual pure states of the ensemble. For two-state systems, the distribution on the surface of the sphere is rotated about the axis determined by the Hamiltonian, as illustrated in ®gure 3.3.

Figure 3.2 Two unravellings of a density operator into pure state projectors. The density operator q on the axis joining the north and south poles is a linear combination of the pure state projectors at the end of the same straight line.

30

Open quantum systems

Figure 3.3 The SchroÈdinger trajectory of a pure state on the Bloch sphere.

The SchroÈdinger dynamics of density operators is obtained directly from the de®nition (3.6) and the evolution of the projectors given by (3.2) and (3.3). The equations are q_ ˆ ÿ…i=hÿ†‰H; qŠ and

q…t† ˆ exp…ÿiHt= hÿ†q…0† exp…iHt= hÿ†:

…3:8a† …3:8b†

For two-state systems, a density operator moves in a circle in the interior of the Bloch sphere, around the axis through the eigenstates of the Hamiltonian, illustrated in ®gure 3.4. The SchroÈdinger evolution is deterministic. The stochastic evolution of quantum systems is more complicated. The pure states perform a kind of Brownian motion in the space of pure state projectors, which for two-state systems is the surface of the Bloch sphere. This is quantum state diffusion. The resulting evolution of the density operator, like the classical probability density of the previous chapter, is deter-

3.3 Entanglement

31

Figure 3.4 The SchroÈdinger trajectory of a density operator in the interior of the Bloch sphere.

ministic, and satis®es a linear differential equation known as the master equation. These processes are the subject of the following chapters. 3.3 Entanglement Entanglement is one of the most puzzling properties of quantum systems. The concept was introduced by SchroÈdinger in 1935 [131, 132, 133, 154]. His briefest de®nition of entanglement was `the whole is in a de®nite state whereas the parts are not'. In modern work on Bell inequalities the word is sometimes used in a narrower sense, but the original broad sense will be used here. For open systems, the system together with its environment is in a de®nite pure quantum state, but the system alone is not. Einstein himself had dif®culty with entanglement, perhaps to the point of disbelief [47, 154]. But it is important to understand it, because it lies behind the quantum theory of open systems and also many of the most fascinating recent quantum experiments, including tests of the Bell inequalities, quantum cryptography, quantum teleportation and proposals for quantum computers.

32

Open quantum systems

In an ensemble, the joint probability distribution PrAB of a combined classical system AB is not usually given by the separate distributions PrA of A and PrB of B, because the states of A and B may be correlated as a result of a past direct or indirect interaction. Correlation is a property of ensembles of systems, not of individual systems. But if two individual classical systems A and B are combined to make a system AB, the state of AB is given completely by the state of A together with the state of B. Remarkably this is not true in quantum mechanics. An individual combined quantum system AB has properties that A and B alone do not. When this happens the two systems A and B are entangled. Systems become entangled as a result of their past direct or indirect interaction. Systems that have not interacted, either directly, or indirectly through other systems, cannot be entangled. If a quantum system AB is in an entangled pure state with projector PAB , then neither A nor B behaves like a pure state. Separately, when A or B interacts with another system C, or when it is measured, it behaves like an ensemble of pure states, represented by a probability distribution over pure states, or by a density operator. These density operators are called reduced density operators of the combined system AB. Let TrA represent the trace over the space of the states of A, and similarly for TrB . Then the reduced density operators for A and for B are qA ˆ TrB qAB

and

qB ˆ TrA qAB :

…3:9†

One of the simplest examples of entanglement is for a spin-zero system that decays into two spin-half systems A and B, which then move apart. The spin states of A and B are then entangled, with pure spin state p jAB i ˆ …1= 2†…jA‡ijBÿi ÿ jAÿijB‡i†:

…3:10†

If we consider A as the system and B as the environment, the density operator of A is given by qA ˆ TrB qAB ˆ TrB jAB ihAB j ˆ 12 IA ;

…3:11†

where IA is the identity in the spin-space of A. The density operator 12 IA is at the centre of the Bloch sphere, representing the spherically symmetric spin distribution. A measurement of the spin of system A or system B alone gives this boring result. But simultaneous measurements of the spins of A and B in different directions tell us about the Bohm form of the Einstein-PodolskyRosen thought experiment, the Bell inequalities, and subsequent laboratory

3.3 Entanglement

33

experiments. This is an example of entanglement between particles. The measurements show a type of correlation of the spins that cannot be produced by classical means. Entanglement in a pure state is not an actual correlation, but measurement of entangled pure states produces a correlated ensemble, so entanglement is potential correlation of a special type. Systems are not always de®ned with ®xed numbers of particles. Even in classical statistical mechanics the number of particles is sometimes allowed to vary, and this is essential for any quantum ®eld theory. A system can be de®ned as the state of a region of space, in which there may be a varying number of particles. An example of entanglement of systems de®ned in this way is quantum interference, as illustrated by the two slits experiment for individual photons shown in ®gure 3.5. We suppose a photon is emitted in a short pulse. The systems A and B are the contents of the regions illustrated when the pulse has just gone through the slits. At this time the system A behaves just like an ensemble with probability 12 of containing one photon, and probability 12 of containing zero photons. The same applies to B. A measurement of the photon number in A and in B is a measurement of the state of the combined system AB. Of the four possible number states of AB, only the two with one photon are seen, so that the measurement shows a correlation between A and B. Before the measurement, this is a potential correlation of the systems A and B, so A and B are entangled. Later on, AB shows the well-known interference properties, which are not obtained by combining the density opera-

Figure 3.5 A two slits experiment with individual photons.

34

Open quantum systems

Figure 3.6 A sketch of a Mach-Zehnder atom interferometer.

tors for A and B separately. It will not show these properties if the number of photons in A or B is measured, even if the result of the measurement is that no photon is seen. The measurement destroys the entanglement. The same behaviour is seen in the Mach-Zehnder atom interferometer illustrated in ®gure 3.6. Readers are warned that some physicists who work on Bell inequalities do not use `entanglement' for this case. Section 8.6 has more on atom interferometers. This illustrates the general properties of entanglement, which only appear in correlations between the results of measurements on A and B, or when A and B are brought together again, as in an interference experiment. Even then the entanglement is often dif®cult to detect, if meanwhile either A or B has interacted with its environment. Entanglement is extremely sensitive to such interactions, which are studied using the theory of open quantum systems, including quantum state diffusion. So although entanglement is a universal property of quantum systems, it can be dif®cult to study experimentally or to use in technology, as for the proposed quantum computers. In general if A and B are entangled, and then stop interacting, the entanglement remains, and it often happens that A and B stay separated, in which case the subsequent separate evolution of A or of B depends on the sum of their Hamiltonian operators. The trace over B of equation (3.8a) or (3.8b) for qAB then shows that the evolution of A is given by equations (3.8) with H ˆ HA and q ˆ qA . So its reduced density operator evolves exactly like an isolated ensemble of identical systems with the same density operator qA , and experiments on A alone cannot distinguish the two. The same applies

3.4 Open systems

35

to B. But simultaneous measurements on A and B can give results that are completely different from measurements on the two isolated and unentangled ensembles. This is the key to quantum nonlocality properties in general and experiments on Bell's inequalities in particular. 3.4 Open systems An open classical system is a system that interacts with its environment, thus producing correlation between them. An open quantum system A is de®ned here as a system that interacts with its environment B, producing entanglement of A and B. But the theory of open systems is mainly concerned with the properties of the system A. These depend on past interactions with the environment, which produce entanglement, but entanglement as such is not usually a major concern of the theory. Typically the environment is much more complicated than the system, and it is impractical to treat the combined system AB. An example is an atom in a gas or in a radiation ®eld, like an atom near the surface of a star. Another is the radiation ®eld of a gas laser, whose environment is a gas of atoms with inverted populations. Another is a resistor in a quantum circuit, in which the environment consists of the internal freedoms of the resistor. Another is a protein in the messy environment of a bacterium. Only in some of these systems is the environment B spatially separated from the system A. In some cases the environment remains in approximate thermal equilibrium at a constant temperature, when it is often called a heat bath. In other cases, as for example the interaction of atoms with a radiation ®eld, the combined system AB may be divided in different ways into system and environment. The internal states of an atom may be the system, and the radiation ®eld may be the environment, as in matter interferometry. Or a few modes of the radiation ®eld may be the system and the internal states of the atoms may be the environment, as for a gas laser. The system may be an atom, and the environment may be some measuring apparatus, as when a state of the atom is measured. Let A be a system and B its environment. The outer boundary of B can be chosen so large that, as far as A is concerned, the interaction of AB with its environment can be neglected. Then AB can be taken to evolve by SchroÈdinger evolution, and the evolution of the density operator qA is obtained by taking the trace: qA …t† ˆ TrB exp…ÿiHt= hÿ†qAB …t† exp…iHt= hÿ†:

…3:12†

36

Open quantum systems

Use 1 and 2 to label two ensembles of the combined system AB, with density operators qAB1 and qAB2 . At time t ˆ 0 a new ensemble of systems AB is made of a sample from ensemble 1 with probability Pr…1† and a sample from ensemble 2 with probability Pr…2†. Then its density operator is qAB …0† ˆ Pr…1†qAB1 …0† ‡ Pr…2†qAB2 …0†:

…3:13†

By the linearity of SchroÈdinger evolution, the same linear combination then applies for all times t. By taking the trace over B, we have that when then

qA …0† ˆ Pr…1†qA1 …0† ‡ Pr…2†qA2 …0†; qA …t† ˆ Pr…1†qA1 …t† ‡ Pr…2†qA2 …t†

(linear evolution). …3:14†

So, for this case, the evolution of the density operator of A is also linear, even though it does not evolve according to SchroÈdinger. Almost always, in practice, the environment is so big and complicated that we cannot solve for AB in order to get the evolution of A, and so we have to approximate. As would be expected, every one of these approximations gives a linear evolution for the density operator qA . In section 7.7 we see that the linearity of density operator evolution is universal and fundamental; it does not depend on any special assumptions or approximations. By contrast, the next chapter on quantum state diffusion shows that the evolution of a state vector for an open system is usually nonlinear. 3.5 Measurement and preparation We start with a theory of quantum measurement which is not very far from the usual one, and later develop it in stages into something very different. Measurement is a physical process by which the state of a quantum system in¯uences the value of a classical variable. The meaning of quantum measurement is here extended to any such process, including laboratory measurements, but also other, very different, processes. Laboratory measurements include the usual particle states that produce the bubbles and sparks of experimental high-energy physics, photon states that produce silver grains in photographic emulsions, and excited states of atoms that produce small currents in solid-state detectors. Other measurements include the photons that send impulses through the optic nerves of owls, the cosmic rays that produced small but permanent

3.5 Measurement and preparation

37

dislocations in mineral crystals during the Jurassic era, and the quantum ¯uctuations in the early universe that are believed to have caused today's anisotropies in the universal background radiation and in galactic clusters. It includes those quantum ¯uctuations that are ampli®ed by chaotic dynamics to produce signi®cant changes in classical dynamical variables. For such measurements there is no measuring apparatus in the usual sense. In all cases, the system that does the measuring is called the measurer. A measurement takes time, because a quantum system takes time to in¯uence a classical variable, but in the elementary quantum theory of measurement this time is ignored. If it is treated as a physical process at all, measurement is supposed to be instantaneous. Measurement that lasts a ®nite time appears in the next chapter. In this chapter we take it to be instantaneous. A measurement on an ensemble of identical pure states of a quantum system typically produces a probability distribution over both the quantum states of the system and the classical states of the measurer. This distribution is conditional on the measurement being made, and so the probabilities are conditional or potential probabilities. Considering them as actual probabilities before any measurement is made leads to confusion, because in quantum mechanics there are incompatible measurements. A measurement is an interaction between the quantum system A that is being measured and the measurer B, which can be considered as the environment of A. Other parts of the environment may interact with the system during the measurement, which experimenters try to avoid in laboratory measurements. We can assume here that these interactions are negligible. In measurement theory, as in any theory of open quantum systems, there is always a problem as to where to put the boundary between the system and its environment. This problem is discussed in section 3.6. Here the system is what is being measured and the environment is the measurer. Let j i be an arbitrary state of a quantum system in an ensemble of identical states with density operator q ˆ P . Let G be the Hermitian operator for one of its dynamical variables, where for simplicity the spectrum of G is discrete and nondegenerate with eigenvalues g, eigenstates jgi and projectors Pg ˆ jgihgj. Unless j i is itself an eigenstate of G, a measurement of g produces an ensemble of different eigenstates of the quantum system, with corresponding values of some classical variable of the measuring system. The probability of making the transition to a particular state jgi is Pr…g† ˆ jhgj ij2 ˆ Tr Pg P :

…3:15†

38

Open quantum systems

During a measurement, an individual system evolves stochastically. It jumps to state jgi with probability Pr…g†. The new state and projector are j 0 i ˆ jgi ˆ hgj iÿ1 Pg j i;

…3:16a†

P0 ˆ Pg ˆ Pr…g†ÿ1 Pg P Pg :

…3:16b†

Because of the normalization factors hgj iÿ1 and Pr…g†ÿ1 , this is nonlinear in j i and nonlinear in P , so it cannot be represented by any kind of SchroÈdinger equation, stochastic or otherwise. This jump process is a short time limit of the quantum state diffusion representation of measurement, as illustrated in the next chapter and demonstrated for a special case in section 5.3. On the other hand, the new density operator obtained as a result of the measurement of the ensemble of identical pure states j i with initial density operator q ˆ P is q0 ˆ

X g

Pr…g†Pg ˆ

X g

Pg qPg ;

…3:17†

which is a linear function of the original density operator q. The ®nal state q0 of the density operator is obtained by taking a mean over the ®nal projectors of (3.16), with weights given by the probabilities. The nonlinearities cancel on taking the mean. This sudden transition from q to q0 is the short-time limit of a linear master equation which represents the interaction of the measured system with the measurer. Equation (3.16a) for a state and (3.16b) for the corresponding projector represent what happens in a single measurement. The density operator equation (3.17) represents what happens in the ensemble of all measurements. The contrast between the nonlinear evolution of individual states represented by the state vector j i or the projector P and the linear evolution of the ensemble represented by the density operator q was emphasized by Gisin [68], and is treated in more detail in chapter 7. Although the ®nal density operator can be obtained uniquely from the ®nal states and their probabilities, the ®nal states and probabilities cannot be obtained uniquely from the ®nal density operator. Although the ®nal density operator q0 for the system is diagonal in g, its unravelling into pure states is not unique, and need not consist of eigenstates of G. For a rz measurement on a two-state system, a possible q0 and a single state and two possible unravellings are shown in ®gure 3.2. In a measurement, the ®nal states of

3.5 Measurement and preparation

39

the ensemble are all eigenstates of G. But one of the unravellings of ®gure 3.2 is not into eigenstates. The correct unravelling is not given by the density operator alone, but needs further conditions. Thus in a measurement, the ®nal density operator does not give full information about individual states of the ensemble. Quantum state diffusion for pure states and the master equation for density operators have similar properties, but their evolution is continuous and takes a ®nite time. Suppose that an initial ensemble of systems is made up of a mixture of pure states j j i with probabilities Pr…j†. Then the initial density operator is qˆ

X j

Pr… j†j j ih j j ˆ

X j

Pr…j†Pj :

…3:18†

The probability that a G measurement will give the result g is then Pr…g† ˆ

X j

Pr… j†Tr Pj Pg ˆ Tr ‰…

X j

Pr… j†Pj †Pg Š ˆ Tr ‰qPg Š

…3:19†

and the density operator after the measurement is q0 ˆ

X g

Pr…g†Pg ˆ

X g

Tr ‰qPg ŠPg ;

…3:20†

which is diagonal in G-representation. Both Pr…g† and q0 depend linearly on the initial density operator q, but do not depend on the way that the density operator is unravelled into pure states with their associated probabilities Pr… j†. Ensembles with the same density operator give the same results for all measurements. The measurements may be delayed, by allowing the ensemble to evolve for a time before making a measurement, either with or without interaction with another system. All such evolutions are linear, with the same result: different ensembles with the same density operator cannot be distinguished by a measurement. That is why this representation is so often used instead of the probability distribution Pr… j†. However, if a system A is entangled with another system B, then measurements on the combined system AB can distinguish between different ensembles of A with the same density operator. The preparation of a quantum system is a physical process by which a classical variable in¯uences a quantum state. The de®nition is used for any such physical process, whether it occurs in the laboratory or not. In a sense, it

40

Open quantum systems

is the inverse of the measurement process. Section 7.5 shows that there are severe constraints on preparation, due to limitations on the number of distinguishable states of physical systems that are usually treated as classical. 3.6 The boundary problem The systems of celestial dynamics are usually treated as if they were isolated. When Newton wanted to understand the dynamics of the motion of the Earth (or Earth and Moon) around the Sun, he ®rst ignored the other planets, but recognized that this was an approximation, which was improved much later by including them. The `system' then took in more than just the Earth and the Sun. The boundary between system and environment was moved outwards to include more of the environment, to improve the approximation. The same holds for open quantum systems. The system of interest may be the state of excitation of an atom. When the atom is in a cavity, the atom can be treated as the system and the radiation ®eld of the cavity as the environment, leading to exponential radiative decay of excited states. But when the atom is in a highly excited Rydberg state and the cavity is of very high quality, some atomic states may resonate with a mode of the ®eld in the cavity, and the original model is no longer good enough. Modes of the radiation ®eld in the cavity have to be included as part of the system, and the walls of the cavity become the most important part of the environment. The choice of boundary between system and environment is the boundary problem. The atom in a cavity is a practical example. For open quantum systems, entanglement between system and environment makes the boundary problem more subtle and more dif®cult than for isolated quantum systems or open classical systems. This is particularly true for the application of quantum state diffusion to measurement, for which the measurer is part of the environment of the system. Where should we put the boundary between the system and the measurer? In the quantum state diffusion theory of measurement this affects our picture of the measurement process. Suppose it is a laboratory measurement of the state of an atom, using emitted photons which are detected using a photomultiplier and counter. The system could be the atom, or it could include part of the photomultiplier, starting with the surface which emits a photoelectron and the electrons which are emitted. When it is just the atom, the measurement is seen as a diffusion process of the state of the atom on the time scale of the decay. But as soon as a part of the photomultiplier is

3.7 Quantum expectation and quantum variance

41

included, the diffusion begins to look like a sudden jump on a far shorter time-scale, as illustrated by various models in the following chapters. In the practical problems of chapter 6, the ambiguity in the boundary between system and environment can be turned to advantage by a careful choice which includes just enough of the environment to provide the precision required and no more. The boundary problem reappears in the foundations of quantum mechanics as discussed in detail in chapter 7. 3.7 Quantum expectation and quantum variance The mean value of a dynamical variable g, in the ensemble formed by measuring g in an ensemble of identical pure states j i, is the quantum expectation or expectation hGi ˆ hGi ˆ h jGj i;

…3:21†

where G is the corresponding operator. The expectation is de®ned in terms of the operator and initial state before measurement, and is therefore a property of the original pure state. In QSD, both actual means over ensembles, denoted by M, and quantum expectations h:i for pure states, which are potential means over ensembles, are important. Suppose that the current pure state of the system is j i and that G is an operator. The quantum expectation of G for the state j i is de®ned by (3.21). This de®nition will be used even when G is not Hermitian, as in the case of creation and annihilation operators. The corresponding shifted operator, which depends on the current state j i and has zero expectation for the current state, is used a lot in QSD theory. It is G
ˆ G ÿ h jGj i ˆ G ÿ hGi:

…3:22†

Commutators of shifted and unshifted operators are the same: ‰G; FŠ ˆ ‰G
; FŠ:

…3:23†

A general operator G may be separated into its real and imaginary (strictly Hermitian and anti-Hermitian) parts GR and iGI , given by GR ˆ 12…G ‡ Gy †

iGI ˆ 12…G ÿ Gy †:

…3:24†

42

Open quantum systems

For the usual annihilation operator a of an oscillator, these parts are given by x=…2 hÿ†2 ; 1

iy=…2 hÿ†2 ; 1

…3:25†

where x and y are conjugate dynamical variables and a; ay ; x and y satisfy the standard commutation relations, ‰a; ay Š ˆ 1;

‰x; yŠ ˆ i hÿ:

…3:26†

Other useful properties of operators and pure states are the quantum variance, which measures the potential variability of an operator G for a pure state j i, and the quantum covariance, which is a bilinear measure of the entanglement, or potential ensemble covariance of two operators F and G. This is the quantum analogue of the classical covariance, which is a bilinear measure of correlation, and has the same advantages and disadvantages. The quantum variance and covariance are de®ned in terms of expectations by analogy with the de®nitions of classical variance and covariance. They can be de®ned for both Hermitian and non-Hermitian operators. For an Hermitian operator G the quantum variance  2 …G† for pure state j i is the classical variance of the variable g in the ensemble obtained by measuring G and G2 . It is  2 …G† ˆ hG2 i ÿ hGi2 ˆ …
g†2

(Hermitian);

…3:27†

where
g is the corresponding standard deviation. The quantum covariance of Hermitian F and G is …F; G† ˆ hFGi ÿ hFihGi

(Hermitian):

…3:28†

If F and G do not commute, then in general …F; G† 6ˆ …G; F†. The quantum covariance of G with itself is just the quantum variance of G. In chapter 5 the de®nitions of quantum variance and covariance are extended to nonHermitian operators. Quantum expectation, variance and covariance are potential statistical properties of a pure quantum state, not actual statistical properties. This is an important distinction, particularly when interpreting experiments where systems are entangled. In QSD, ensemble means and variances for ensembles of quantum systems have to be distinguished from the quantum expectations and quantum variances of the individual states.

3.7 Quantum expectation and quantum variance

43

An open quantum system continually interacts with its environment. Its `mixed state' is traditionally represented by a density operator q. In QSD, an individual quantum system is always considered as a pure state represented by a state vector j i, whether it is isolated or open, whether it is interacting with its environment or not. The density operator is used, but plays a secondary role.

4 Quantum state diffusion 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8

Master equations QSD equations from master equations Examples Projectors Linear unravelling Other ¯uctuations QSD, jumps and Newtonian dynamics The circuit analogy

Chapter 2 showed how Brownian motion, like the dynamics of all open classical systems, can be represented either as the solution of a stochastic equation for an individual system of the ensemble, or in terms of a probability distribution that satis®es a linear deterministic equation. Chapter 3 introduced open quantum systems from the viewpoint of probability distributions and the density operator. This chapter presents the dynamics of open quantum systems, starting in the ®rst section with the linear deterministic master equation for the density operator, and then using it to derive the stochastic quantum state diffusion equations for an individual system of the ensemble. These QSD equations for state vectors are (4.19), (4.20) and (4.21). Measurement and dissipation in two-state systems illustrate the theory by example. Sections 4.4±4.7 present QSD equations for projectors and also alternative stochastic equations that do not satisfy the conditions of QSD. Much of this chapter and the next follows the papers of Gisin and Percival [72, 73, 74, 111].

4.1 Master equations Interaction of an open quantum system with the environment produces stochastic changes that can be represented by an ensemble of changing pure states j …t†i with probability distribution Pr… …t††, or by a changing density operator q…t†. The density operator satis®es a linear master equation. Lindblad [91] showed that if any physical density operator q…t† satis®es a linear differential equation, then this equation can always be put into the Lindblad form 44

4.1 Master equations

 X i q_ ˆ ÿ ÿ ‰H; qŠ ‡ Lj qLyj ÿ 12 Lyj Lj q ÿ 12 qLyj Lj h j

45

(Lindblad),

…4:1†

where the Lindblad operators or Lindblads Lj may or may not be Hermitian. Master equations have been obtained for a wide variety of open systems, thereby de®ning Lindblads for these systems. Only the Hamiltonian and Lindblads are needed to formulate QSD for a particular system, and so the experience gained in formulating master equations for open systems can be applied directly to QSD. Measurement is a special case, which is treated in detail later. Just as the Hamiltonian operator H determines the contribution of the internal deterministic dynamics to the change of q, so the Lindblad operators Lj determine the contribution of the stochastic dynamics due to interaction with the environment. Later we will see, for example, that an Hermitian Lindblad operator can be used to represent the effect of a measurement, and that a non-Hermitian annihilation Lindblad operator represents a dissipation. In the isolation limit, when the effect of the environment is negligible, the Lindblads may be neglected, the Hamiltonian remains, and each system of the ensemble evolves by SchroÈdinger's equation. In the opposite limit, when the effect of the Hamiltonian is negligible and the effect of the environment dominates, the systems are wide open, and their evolution is determined by the Lindblad terms. The theory of wide open systems is relatively simple and has many applications. The same theory can often be applied to systems in which the Hamiltonian is important, by using a form of interaction representation to remove the Hamiltonian from the equations. The simplest type of wide open system is the elementary open system with only one Lindblad and with master equation _ ˆ Lq…t†Ly ÿ 12 Ly Lq…t† ÿ 12 q…t†Ly L: q…t†

…4:2†

There are many types of interaction with the environment. We illustrate the properties of master equations and QSD equations using elementary open systems with simple Lindblads representing measurement and dissipation. In this chapter we take account of the fact that measurement takes a ®nite time. The measurement of a dynamical variable with Hermitian operator G is represented by the Lindblad L ˆ cG, where c is a constant that determines the rate of measurement. Using jgi-representation, the solution of the master equation for g0 g …t† with initial values g0 g …0† is

46

Quantum state diffusion

g0 g …t† ˆ exp‰ÿc2 …g0 ÿ g†2 tŠg0 g …0†:

…4:3†

The off-diagonal elements are suppressed exponentially in time at a rate which increases with the difference between the eigenvalues. When c is large, the measurement takes place rapidly, and tends to become a sudden jump as described in the previous chapter. For a two-state system, the Bloch vector r~…t† for a rz measurement with an initial state given by r~…0† is r~…t† ˆ …exp…ÿc2 t†x…0†; exp…ÿc2 t†y…0†; z…0††:

…4:4†

The measurement produces an exponential drift along a straight line directly towards the axis joining j‡i and jÿi states at the north and south poles of the Bloch sphere. A typical dissipative process is the radiative decay of the excited state of a two-state atom. The excited state is jei ˆ j‡i, the ground state is jgi ˆ jÿi. The decay process is represented by the operator 1 2

1 2

rÿ ˆ …rx ÿ iry † ˆ 1 2

1 2



0 1

 0 ; 0

…4:5†

1

whose Hermitian conjugate is 2 r‡ . is the rate of decay. Also r‡ rÿ ˆ P‡ , the projector onto the excited state. The elementary Lindblad master equation is q_ ˆ rÿ q r‡ ÿ 12 …P‡ q ‡ qP‡ †:

…4:6†

The general initial condition is a linear combination of the ground and excited states, with Bloch vector r~…0†. The time dependence of the density matrix is 

‡‡ …0† eÿ t ÿ‡ …0† eÿ t=2

 ‡ÿ …0† eÿ t=2 : ÿÿ …0† …1 ÿ eÿ t †

…4:7†

The Bloch vector is given by   r~…t† ˆ x…0† eÿ t=2 ; y…0† eÿ t=2 ; …z…0† ‡ 1† eÿ t ÿ 1

…4:8†

and moves along a parabolic path to the south pole, which represents the ground state.

4.2 QSD equations from master equations

47

The master equation for an elementary system is invariant under multiplication of L by a complex phase factor u of unit modulus. The same effect is produced by L and by uL, even when L is Hermitian and uL is not. For elementary systems, one Lindblad equation uniquely represents the evolution of the ensemble. The Lindblad equation for a general system is similarly invariant for multi~ of Lindblads Lj by a unitary matrix ukj . The new plication of the vector L P equation, obtained by replacing the Lj by L0k ˆ ukj Lj is also of Lindblad form. It may be considered as a different but equivalent equation, but it is preferable to consider it as the same equation with a different representation for the Lindblad vector in a linear space. The QSD equations also have this unitary invariance property, which is not shared by all stochastic diffusion equations for quantum states. The ambiguity in the choice of Lindblads is a nuisance, so it is helpful to adopt a convention which reduces the choice. The convention is that where possible Hermitian Lindblads are used in preference to others, giving a standard form for this case. 4.2 QSD equations from master equations An ensemble of quantum systems, whose state vectors satisfy a stochastic differential equation, has a density operator that satis®es a unique deterministic differential master equation. By contrast a master equation corresponds to many different stochastic equations for its component states. These are known as different unravellings of the master equation. QSD is one such unravelling. This section shows how the quantum state diffusion equation is derived from the master equation (4.1) for an elementary system. A density operator q for an evolving quantum system can be expressed in many ways as a mean M over a distribution of normalized pure state projection operators. We seek differential equations for j …t†i such that the density operator given by the ensemble mean over the projectors q…t† ˆ M j …t†ih …t†j

…4:9†

satis®es the Lindblad master equation (4.1). In time dt the variation jd i in j i is then given by the Itoà equation jd i ˆ jdriftidt ‡ jfluctid ˆ jvidt ‡ jf id:

…4:10†

48

Quantum state diffusion

The evolution of the state vector is continuous. The differential stochastic ¯uctuation d has equal and independent ¯uctuations in its real and imaginary parts (chapter 2): M d ˆ 0;

M …d†2 ˆ 0;

M jdj2 ˆ dt:

…4:11†

Fluctuations at different times are assumed to be statistically independent, so the stochastic differential equation represents a Markov process. The state vector remains normalized. These four conditions lead to a unique unravelling into states whose trajectories in Hilbert space satisfy the QSD equations. The other unravellings described in sections 4.5 and 4.6 fail to satisfy at least one of these conditions. First we derive the dominant ¯uctuation terms, and then the drift. To preserve normalization, keeping the state vector on the unit sphere, the differential change in the state vector due to the ¯uctuation must be orthogonal to the state vector itself, because the ¯uctuations are of higher order than the drift. So h jf i ˆ 0:

…4:12†

Taking means over jd i and jd ihd j in equation (4.10), M jd i ˆ jvidt



M jd ihd j ˆ jf ihf jdt;

…4:13†

so the change in the density operator is given by

or

dq ˆ M …j ihd j ‡ jd ih j ‡ jd ihd j† q_ ˆ …j ihvj ‡ jvih j ‡ jf ihf j†:

…4:14†

The stochastic terms are determined by the component of q_ in the space orthogonal to j i, so, using the master equation (4.1), jf ihf j ˆ …I ÿ j ih j† q_ …I ÿ j ih j† ˆ …I ÿ j ih j† LqLy …I ÿ j ih j† ˆ …I ÿ j ih j† Lj ih j Ly …I ÿ j ih j† ˆ j…L ÿ hLi† ih…L ÿ hLi† j ˆ jL
ihL
j; where hLi and L
are de®ned by (3.21) and (3.22). So

…4:15†

4.2 QSD equations from master equations

jf i ˆ u…L ÿ hLi†j i;

49

…4:16†

where u is a phase factor for the operator L. However in the QSD equation (4.10) the distribution of the ¯uctuations d given by (4.11) is invariant under multiplication by a phase factor, so that u makes no difference to the stochastic equations and can be omitted. The drift is given by _ i ˆ j ihvj i ‡ jvi; qj

…4:17†

from which it follows on premultiplying by h j that _ i Reh jvi ˆ 12h jqj _ ij i ‡ i…t†j i _ i ÿ 12h jqj so that jvi ˆ qj   ˆ hLy iL ÿ 12Ly L ÿ 12hLy ihLi ‡ i…t† j i;

…4:18†

where the last line follows from the master equation (4.1) with q ˆ j ih j. The pure imaginary function of time i…t† changes only the phase factor of the state vector, and so is nonphysical. It is chosen by convention to be zero, so that the phase change agrees with the SchroÈdinger equation in the absence of interaction with the environment. So the elementary QSD equation is   ) jd i ˆ hLy iL ÿ 12Ly L ÿ 12hLy ihLi j idt ‡ …L ÿ hLi†j id   ˆ hLy iL ÿ 12Ly L ÿ 12hLy ihLi j idt ‡ L
j id:

…4:19†

For Hermitian L this becomes jd i ˆ ÿ12 L2
j idt ‡ L
j id;

…4:20†

where L
ˆ L ÿ hLi is the shifted L-operator whose expectation for the current state j i is zero. This QSD equation represents the measurement of a dynamical variable whose operator is proportional to L. It can be shown that the mean of the expectation M hLi remains constant, as would be expected from a measurement. For general systems, the derivation of the QSD equation follows the same lines, with a unitary transformation of the Lj instead of the phase factor u in

50

Quantum state diffusion

(4.16). Given the Lindblad master equation (4.1) for the density operator q, the stochastic differential equation for the state vector j i is  X y i hLj iLj ÿ 12Lyj Lj ÿ 12hLyj ihLj i j idt jd i ˆ ÿ ÿ Hj idt ‡ h j Xÿ
Lj ÿ hLj i j idj : ‡

…4:21†

j

The general equation follows from SchroÈdinger's equation and the equation (4.19) for an elementary system on using the additivity rule of section 2.9. Because of the rule, this book concentrates throughout on elementary systems with no Hamiltonian and only one Lindblad. The ®rst sum in (4.21) represents the drift of the state vector and the second sum the independent random ¯uctuations due to the different interactions of the system with its environment. hLj i ˆ h jLj j i is the expectation of Lj for state j i, and the density operator is given by the mean over the projectors onto the quantum states of the ensemble: q ˆ M j ih j;

…4:22†

where M represents a mean over the ensemble. The dj are independent normalized complex differential random variables, satisfying the complex relations from (2.33), which are M dj ˆ 0; M …dj dj 0 † ˆ 0;

M …dj dj 0 † ˆ jj 0 dt:

…4:23a† …4:23b†

Unlike the master equation, the QSD equation (4.21) is nonlinear, because the scalars hLj i depend quadratically on j i. If the time dependence of the scalars is known, then the state can be obtained by solving a linear stochastic differential equation, suggesting a possible iterative solution. The QSD equations are also usually nonlocal, which means that the solutions depend on distant in¯uences. This can be seen for the case of the elementary open system whose state is represented by a wave function …x; t†. Suppose the Lindblad in the QSD equation (4.19) is the position operator L ˆ x. The valueR of …x0 ; t ‡ dt† for a particular point x0 depends on the expectation hxi ˆ dx
j …x†j2 x, which depends on all the non-zero

4.3 Examples

51

Figure 4.1a A single trajectory on the Bloch sphere for H ˆ 0; L ˆ cz :

values of …x; t†, however far from x0 the point x may be. This nonlocality causes problems for relativistic QSD, as discussed in section 8.7. The solution of a QSD equation is a nondifferentiable function of time. Its Fourier transform is like the Fourier transform of white noise, whose amplitudes are signi®cant up to arbitrarily high frequencies. Bohr's relation E ˆ hÿ! implies correspondingly high energies, but there are no such energies, so the usual Fourier relation between energy and time breaks down in nonrelativistic QSD. But the Fourier relation between momentum and position is retained, introducing a fundamental asymmetry that is incompatible with special relativity. 4.3 Examples Consider the examples of two-state measurement and dissipation. In each case the QSD projector diffuses on the surface of the Bloch sphere. For a rz measurement, the diffusion is illustrated in ®gure 4.1a. Figure 4.1b illustrates the time dependence of hrz i for the same trajectory.

52

Quantum state diffusion

Figure 4.1b The time dependence of hrz i:

The projector diffuses around the surface of the sphere to either the north or south pole, corresponding to the j‡i or jÿi state of the system. The probability of diffusion to each of these states from an initial state j …0†i is given by initial expectations of the corresponding projectors. This is an illustration of the expectations for states becoming probabilities for the ensemble after the measurement. This is shown for a small sample in ®gure 4.2. Notice that for QSD there is no ambiguity in the unravelling. The ensemble of trajectories is de®ned uniquely. The equations not only give the correct evolution of the density operator, which was an input into their derivation, but they also give the unravelling of the density operator into evolving pure states, all of which approach the eigenstates j‡i and jÿi. This long-term behaviour of individual systems was not put in as a condition at any stage. Of course it is not possible to predict which of the trajectories of the ensemble will be seen in a given experiment. That depends on the ¯uctuation, which is unknown. It may seem surprising that diffusing states concentrate at two points on the sphere. It is as if the sphere becomes more and more sticky as the state

4.3 Examples

53

Figure 4.2 The time dependence of hP1 i for seven trajectories with the same initial p condition j …0†i ˆ …1= 2†…j‡i ‡ jÿi†, but different ¯uctuations.

approaches an eigenstate, so that the states of the system ®nd it more dif®cult to diffuse away than to diffuse towards the eigenstates. Figures 4.3a and 4.3b illustrate the radiative decay of a two-state atom from the excited state. The diffusion typically takes a time comparable to the decay time, with a signi®cant proportion of this time in intermediate states between the excited and ground states. This is not seen in the laboratory, where the emitted photons can be counted, and sudden quantum jumps are seen. Such jumps are also seen in QSD as a result of simultaneous measurement and radiative decay, which is the correct QSD representation of the decay process as seen in the laboratory. The QSD equation is no longer elementary, but has two Lindblad operators, L1 ˆ rÿ

…decay†;

L2 ˆ c rz

…measurement†:

…4:24†

The time dependence of hP1 i is illustrated in ®gures 4.4a±c for ˆ 0:3; c ˆ 0:15; 0:3 and 8. As the measurement Lindblad increases, the continuous

54

Quantum state diffusion

Figure 4.3a Radiative decay: a trajectory on the Bloch sphere.

Figure 4.3b The time dependence of the expectation of the projector hP1 i ˆ 12 …1 ‡ hrz i†, which is ‡1 for the excited state, and 0 for the ground state.

4.3 Examples

Figure 4.4 For caption see next page.

55

56

Quantum state diffusion

Figure 4.4 The time dependence of the expectation of the projector onto an excited state during radiative decay for the same decay constant ˆ 0:3. The measurement constant is c ˆ 0:15, 0.3 and 8.0 for (a), (b) and (c) respectively.

diffusion becomes more and more like a jump. These are not instantaneous jumps, but their sharpness depends on the nature of the interaction of the system with its environment, through L2 . Their mean rate is determined by L1 alone. A comparison of this example with [155], which also uses the Bloch sphere for illustration, highlights the differences between QSD, in which the jumps appear as a limit of continuous diffusion, and the quantum jump approach in which state diffusion appears as a special limit of quantum jumps. This is discussed in more detail in section 4.7. Figures 4.5a±d illustrate a forced and damped linear oscillator in interaction representation, for which the Hamiltonian and Lindblad operators are H ˆ 2i…ay ÿ a†;

Lˆ a

…4:25†

and the initial state is j 0 i ˆ j5i, where j5i is a Fock number state representing the ®fth excited state. For ®gures 4.5a,b, there is zero forcing,  ˆ 0.

4.3 Examples

Figure 4.5 For caption see next page.

57

58

Quantum state diffusion

Figure 4.5 Phase space plots and expectation of the photon number or energy for a forced, damped linear oscillator.

4.3 Examples

59

The ®gures illustrate state diffusion towards a coherent state, which moves like a classical particle and has no stochastic ¯uctuations. This is an example in the quantum domain of localization in phase space, which is characteristic of our everyday experience in the classical domain. We use localization in the sense of the state of an individual system of an ensemble being con®ned to the neighbourhood of a single state or group of states. We use it for a quantum state being localized in the neighbourhood of a point in phase space, as for a con®ned wave packet. We also use it to describe the processes that lead to these localized states, sometimes called reduction in the traditional theory of quantum measurement. The next example is a quantum cascade for an oscillator with emission and measurement, also in interaction representation. The operator L1 represents the interaction required to measure the system in a given state, whereas L2 represents the damping due to photon emission. The operators are H ˆ 0;

L1 ˆ 6ay a;

L2 ˆ 0:1a:

…4:26†

Figure 4.6 shows that the continuous state diffusion approximates to a succession of sudden jumps. Our ®nal example, illustrated in ®gure 4.7, shows how a Hamiltonian can turn momentum localization into phase space localization, with a resultant position localization. This is a very simple example of a general phenomenon for systems of many freedoms, in which the dynamics of the Hamiltonian can produce phase space localization from a few simple Hermitian Lindblads, even just one. The operators are H ˆ 12 …q2 ‡ p2 †;

L ˆ 0:3p:

…4:27†

The initial condition is the Fock number state j5i and the graph shows hqi and hpi as a function of time, with quantum standard deviations, the square root of the variance, for each. The Hamiltonian causes a rotation in phase space, which turns a wave packet that is localized in momentum into a wave packet localized in position, with the result that both of them localize, though not quite to a minimum indeterminacy wave packet for these conjugate variables.

60

Quantum state diffusion

Figure 4.6 Sudden jumps approximated by continuous state diffusion.

Figure 4.7 Momentum localization for an oscillator. The vertical bars are
q and
p. The graph of hpi is distinguished by the continuous line at the centre of the bars.

4.4 Projectors

61

Other examples are given in the papers listed at the end of the chapter summary, including symmetry breaking by localization and a QSD picture of measurement by the formation of a photographic latent image. Spiller and his collaborators have studied thermal equilibrium and the behaviour of quantum capacitors [139, 137]. Further examples are given in more detail in chapter 6.

4.4 Projectors Quantum state diffusion has a particularly elegant representation in terms of projection operators, or projectors [34]. It follows directly from the ¯uctuation term L
j id in the elementary QSD equation (4.19) that the equation of change for the pure state projector has the form dP ˆ d…j ih j† ˆ Xdt ‡ L
P d ‡ P Ly
d

…4:28†

and it follows from the master equation (4.1) for the ensemble with initial density operator P that 

Xdt ˆ M dP ˆ dq ˆ LP Ly dt ÿ 12 …Ly LP ‡ P Ly L†dt;

…4:29†

giving dP ˆ LP Ly dt ÿ 12 …Ly LP ‡ P Ly L†dt ‡ L
P d ‡ P Ly
d

(elementary).

…4:30†

The generalization from the elementary equations for one Lindblad and no Hamiltonian to the general equations with a Hamiltonian and any number of Lindblads is obvious. In this representation the connection between the QSD equation for the state vector and the master equation for the density operator is particularly clear.

62

Quantum state diffusion

4.5 Linear unravelling If we just omit the nonlinear terms from the QSD equations we get the much simpler linear equations jdi ˆ ÿ12Ly Ljidt ‡ Ljid:

…4:31†

The linear form of the projector equations is exactly the same as above (4.30), with L in place of L
, and it is obvious that the mean gives the usual master equations. So why isn't the linear form always used? It is because the norm of the individual state wave functions is not preserved. The change in the norm is given by d…jj2 † ˆ Tr dP ˆ hjLjid ‡ hjLy jid ˆ 2hjLR jidR ÿ 2hjLI jidI :

) …4:32†

For linear unravelling the ensemble is not a normal ensemble in which each member of the ensemble has equal weight, but a weighted ensemble, in which the weight is given by the changing norm. Linear unravelling has many advantages, because the analysis is much simpler. In particular the path integral formulation is relatively straightforward for linear unravelling. But it also has a major disadvantage. The j…t†i of linear unravelling does not represent an individual physical system. Even if each solution of the stochastic equations is normalized, giving j0 …t†i ˆ j…t†i=j…t†j;

…4:33†

the normalized states are not solutions of the ordinary QSD equations, their mean does not give the master equation and they do not represent the evolution of an individual system. Thus a solution of the linear equations cannot be used to represent the evolution of an individual system, such as a single run of a laboratory experiment. A consequence of this is that when the solutions of the linear stochastic equations are used in a Monte Carlo numerical method, the sample tends to be dominated by the single member with the greatest norm, so that the statistics are very bad when the runs last a signi®cant period of time. The statistics can be improved by simulating biological reproduction of the members with large norm, but this complicates the programming, and is still inferior to the usual nonlinear numerical Monte Carlo method described in

4.6 Other fluctuations

63

chapter 6. Another consequence is that the linear stochastic equations do not satisfy Bell's conditions for a good quantum theory in chapter 7. Hudson and Parthasarathy have developed the mathematics of linear unravelling [85, 102], whereas Goetsch and Graham treat the physics [79]. 4.6 Other ¯uctuations We now return to the nonlinear theory. The unitary condition described after equation (4.8) is a consequence of having complex ¯uctuations satisfying the conditions (4.11) or (4.23). We now relax these conditions, allowing other types of ¯uctuation. Since these conditions are needed to obtain a unique QSD equation from a master equation, there is then a large choice of possible quantum state diffusion equations consistent with a given master equation. The two most important can be expressed in terms of real ¯uctuations dw de®ned in section 2.9. They can be illustrated for a two-state system with the Hermitian Lindblad L ˆ rz on the Bloch sphere of ®gure 3.1. First suppose that d is replaced by dw in the ¯uctuation term. Then the state diffuses along a line of longitude to either the north or south pole of the Bloch sphere. Like QSD, this state diffusion process localizes, and represents a measurement of the z-component of the spin. Such state diffusion equations can be used to represent measurement. They have been used extensively by researchers on continuous spontaneous localization, discussed in more detail in chapter 7. Secondly suppose that d is replaced by ÿidw, so that the ¯uctuations are pure imaginary. Then, for the two-state system with Lindblad L ˆ rz , the state diffuses along a line of latitude. There is no localization, and the theory cannot be used to represent a measurement. This form of state diffusion represents a system with a ¯uctuating Hamiltonian, for example a charged particle or atom in a classical electromagnetic ®eld. Ordinary QSD with complex ¯uctuations is used to represent interactions between system and environment, which produces entanglement between them. The pure imaginary ¯uctuations represent the action of a stochastic environment on the system, represented by a ¯uctuating Hamiltonian, which does not produce entanglement. The general form of the state diffusion equations with real and pure imaginary ¯uctuations is the same as for the QSD equations, with d replaced by dw or by idw. The measurement and the ¯uctuation result in the same master equation and the same evolution of the density operator. So the two processes cannot be distinguished from each other or from QSD by looking at the density operator. However, they represent different evolutions for the individual systems. For both real and imaginary ¯uctuations, it must be

64

Quantum state diffusion

remembered that the theory is no longer invariant under unitary transformations in the space of the ¯uctuations, nor in the space of the Lindblad operators. Consequently, phase factors in these operators now make a difference to the evolution of the individual states of the ensemble. 4.7 QSD, jumps and Newtonian dynamics So far we have assumed that the quantum states satisfy a stochastic differential equation with continuous solutions. Not all unravellings have this property. There is also the possibility that the stochastic process consists of a sequence of discrete quantum jumps. Such unravellings are very commonly used for the solution of practical problems in quantum optics, and they have also been used as the foundation of an alternative quantum theory. In comparing QSD and jumps it is useful to consider an analogy. For the Newtonian dynamics of the Earth and the Moon, momentum is continuously transferred between them. But in the impulsive dynamics of billiards, snooker, pool or Newton's balls, momentum is supposed to be transferred instantaneously. A continuous change of momentum can be interpreted as a sequence of small sudden changes or impulses, and this picture is often appropriate for calculations and theory. Conversely, an impulse can be considered as the short-time limit of a continuous transfer of momentum, with the magnitude of the impulse held constant. Either approach is legitimate, but in classical dynamics the continuous theory is usually treated as fundamental and the sudden changes or impulses are considered to be secondary. Similarly in the unravelling of master equations, either a continuous stochastic equation like a QSD equation, or a sequence of sudden jumps, can be considered as fundamental, and the other derived from it. In this book we show how the jumps can be derived as a limit of QSD. In quantum optics the interaction between an electromagnetic ®eld and a photon counter is continuous. But the response of a photon counter to absorption of a photon is very rapid. It can be considered as instantaneous, and it therefore makes sense to model these measurements of quantum optics by jumps, and that is what is usually done in practice. By contrast, where the response of the environment is naturally considered as continuous, for example in the interaction of an organic molecule with its liquid environment, and in the examples given in chapter 6, QSD is more appropriate. States can be localized both with QSD and with jumps [83]. Localization appears to be more natural for QSD. Localization is a physical process, but its details are not accessible to experiment. We have no way of deciding from experiment whether it is a continuous diffusion process or a series of stochastic jumps. Localization can

4.8 The circuit analogy

65

be based on continuous stochastic changes or on quantum jumps as fundamental. Either is legitimate, but here, by analogy with Newtonian dynamics, the continuous theory is treated as fundamental, and the jumps as secondary. It should be remembered, as in the case of hard collisions between solid bodies in Newtonian dynamics, that where sudden changes are more important, as in photon counting, jumps are often simpler in theory and in practice. This is important when deciding whether to use QSD in the computer program described in chapter 6. The application of quantum jump methods to quantum optics was stimulated by Dehmelt's three-level scheme for observing them in single trapped atoms [29, 30]. The method has an extensive literature, including [22, 121, 148, 24, 160, 45, 96, 25, 155, 156]. Garraway and Knight have made comparisons between QSD and quantum jump simulations for dissipative systems in quantum optics [57, 59, 58]. They showed that QSD often localizes where jump simulations do not. With Steinbach they introduced a Monte Carlo method which provides a fourthorder simulation of the master equation [147]. Holland, Marksteiner, Marte and Zoller have shown how to produce localization with jumps by simulating ®ctional measurements [83]. The time correlation functions and spectra of open quantum systems [90] can be evaluated using quantum jump methods [22, 54] and QSD [128], but it is not simple to ®nd a correlation function for a single QSD trajectory with the correct ensemble mean. This problem has been addressed by Gisin using the Heisenberg picture [71], and by Brun and Gisin who introduce a supplementary state vector trajectory that depends on the original trajectory through an additional QSD equation [20]. 4.8 The circuit analogy The representation of measurement by an Hermitian Lindblad is formal. One of the aims of QSD is to show that this formal treatment of measurement follows from a more detailed treatment of the interaction between the system and the apparatus as its environment, but this depends on the details of the physics of the measuring apparatus, and the choice of boundary. It has not yet been worked out in suf®cient detail. There is an analogy between the state diffusion theory of quantum mechanics and classical circuit theory. An electric circuit includes Hamiltonian elements, with capacitance and inductance, and also nonHamiltonian elements, such as resistors, which represent the interaction of the circuit with its `environment', consisting of the internal electronic

66

Quantum state diffusion

freedoms of the resistor. This interaction produces stochastic ¯uctuations in the form of thermal noise, which can be represented by an Itoà equation. The resistors are normally treated on the same basis as the Hamiltonian parts of the system. It is not considered necessary to analyse the detailed physics of every resistor before attempting to solve a problem in circuit theory. But it is important that such an analysis should be possible, for without it there can be no understanding of the physics of the process. A simple example is the aerial treated as a resistor. A more complicated example is the detailed dynamics of electrons in a solid state resistor. This is obviously much more dif®cult than the representation of a linear resistor by a real impedance in circuit theory. The simple representation is the analogue of the representation of a quantum measurement by an Hermitian Lindblad, the electron dynamics is the analogue of the treatment of measuring apparatus as the environment of the quantum system and working out the interactions between them in detail. For example, this would require an analysis of the physics of ampli®ers and electron multipliers. It has not yet been done in general, but a special case appears in section 6.5.

5 Localization 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8

Measurement and classical motion Quantum variance and covariance, ensemble localization Quantum measurement Dissipation Channels and statistical properties Localization theorems Proof of dispersion entropy theorem Discussion

The localization properties of QSD are important for practical numerical computations and essential for its application to quantum foundations. When a physical variable represented by an Hermitian operator is measured, the state localizes towards a single value of that variable. For more general interactions of the system with its environment, the states of the system tend to localize to wave packets that are con®ned to small regions of phase space. These properties have been con®rmed numerically for a wide variety of simple systems. General properties for wide open systems are derived in this chapter from the QSD equations themselves. Localization produced by QSD is so important that the theory is explained in detail in this chapter, which is more mathematical than the other chapters. If you are only interested in the main results, you can ®nd them in the ®rst and last sections. 5.1 Measurement and classical motion Stern and Gerlach passed a beam of silver atoms through a magnetic ®eld and detected them with photographic emulsion. The beam was split into two, corresponding to the components of spin in the direction of the magnetic ®eld, as illustrated in ®gure 5.1. In a Stern-Gerlach experiment, when the spin of an atom in the initial beam is aligned in a direction perpendicular to the ®eld, the de Broglie wave of the atomic motion is a linear combination of waves of equal amplitude in each beam. The probabilities are equal too, but a single atom is only seen in one of them. The process of detection, which produces a microscopic permanent image of each atom on a photographic plate, appears to localize the atom in one beam or the other. This is sometimes called reduction. The 67

68

Localization

Figure 5.1 Stern-Gerlach apparatus and localization of silver atoms by a photographic emulsion.

Stern-Gerlach experiment is typical of many quantum measurements in the laboratory. When a gamma ray is emitted by a nucleus, its wave is spread around the nucleus in all directions, but when it is detected by absorption in a ¯uorescent screen, its effect is seen as a localized ¯ash of light. This is the old problem of wave-particle duality, seen in ®gure 5.2. These experiments and others like them cannot be explained by using only SchroÈdinger's quantum dynamics. For Bohr and von Neumann and almost

Figure 5.2 Localization of a gamma ray emitted by a nucleus.

5.1 Measurement and classical motion

69

Figure 5.3 The track of a classical Brownian particle and a sketch of the corresponding outgoing de Broglie wave.

all practising quantum physicists since then, the theory of the experiments requires an interpretation, which is additional to the dynamics. As shown by Ehrenfest, classical motion of free particles can be simulated for short periods of time by localized wave packets that are solutions of SchroÈdinger's equation. Later, according to the equation, these wave packets disperse more or less rapidly into larger and larger regions of space. Yet even small systems like Brownian particles insist on behaving classically for arbitrarily long periods of time, as if there were some process that keeps them localized, as shown in ®gure 5.3. All classical dynamical variables behave as if there were a process that keeps them localized. According to QSD there is a process that keeps all classical dynamical variables localized, and which also produces the localization in quantum measurements. It is quantum state diffusion. This was the reason for introducing QSD in the ®rst place. When the diffusion takes place as a result of an interaction of the system with its environment, it is immaterial whether that environment happens to include measuring apparatus, or some more general environment. Almost any interaction with an environment produces localization, with different environments producing different kinds of localization. Measurement is nothing special. The strengths and weaknesses of QSD and

70

Localization

the usual interpretation of quantum measurements are discussed in more detail in chapter 7. Localization is also the key to computing the evolution of open quantum systems. The computation requires a set of basis states. The number of these needed for a localized wave packet is often many orders of magnitude less than the number needed for an unlocalized wave, making computations possible for systems whose evolution is otherwise uncomputable. Localization can even make the evolution of an open quantum system easier to compute than when the system is isolated. Chapter 6 has a detailed account of the use of localization in QSD computations. In section 5.2, the de®nitions of variance and covariance in section 3.7 are extended to non-Hermitian operators. Sections 5.3 and 5.4 present simple examples, of localization for one dynamical variable, and of localization in phase space. The rest of the chapter describes in more detail the physics of localization, particularly in position space. General properties are summarized in four principles and two theorems. The principles are: Pri1. Interactions with independent parts of the environment lead to statistically independent quantum state diffusions. Pri2. Amplitudes ¯uctuate where probabilities are conserved: where there are independent environmental interactions in subspaces, ensemble probabilities are conserved, but amplitudes for the individual systems of the ensemble ¯uctuate, unless the effective rate of interaction of the system with its environment is zero. Pri3. The above probability conservation and ¯uctuation together imply localization (with special exceptions). Pri4. Position localization of states in position space follows from locality of interactions. The rest frame of the localization is determined by the environment.

The theorems are the dispersion entropy theorem and the general localization theorem. These appeared in [111]. A classical theory of localization is presented in chapter 9, which deals with the localization of systems with high quantum number, whereas a theory of the localized states is given in section 10.6. 5.2 Quantum variance and covariance, ensemble localization The de®nitions of section 3.7 for Hermitian operators are here extended to arbitrary operators. They are natural generalizations of the classical covar-

5.2 Quantum variance and covariance, ensemble localization

71

iance of complex variables in section 2.6, and many of the results of that section carry over to this one. The quantum covariance of two operators for a system in the state j i is …F; G† ˆ hFy
G
i ˆ hFy
Gi ˆ hFy G
i ˆ hFy Gi ÿ hFi hGi:

…5:1†

The quantum covariance of F with itself is just the quantum variance of F: …F; F† ˆ hFy Fi ÿ hFi hFi ˆ  2 …F†:

…5:2†

For Hermitian F,  2 …F† ˆ …
f †2 , where
f is the standard deviation of the measured dynamical variable f . When F is non-Hermitian and it does not commute with Fy , the variance  2 …F† is not the same as that of its conjugate Fy , but their average is the sum of the variances  2 …FR † and  2 …FI †. This convention is helpful to the later analysis, but differs from an earlier convention of Caves [23]. In particular, for annihilation and creation operators,  2 …a† ‡ 12 ˆ  2 …ay † ÿ 12 ˆ


1 ÿ 2  …x† ‡  2 …y† : ÿ 2h

…5:3†

If  2 …a† ˆ 0, then  2 …x† ‡  2 …y† ˆ hÿ;

…5:4†

the value for the ground state of an oscillator, or for the more general coherent states, which are minimum indeterminacy wave packets. The ensemble localization …F† of an Hermitian or non-Hermitian operator F is de®ned as the inverse of the ensemble mean of the quantum variance: ÿ
ÿ1 …F† ˆ M  2 …F† :

…5:5†

According to this de®nition, the localization of the annihilation operator a is in®nite for a minimum indeterminacy coherent state. With these de®nitions there are strict lower bounds on the rates of localization for wide open systems.

72

Localization

5.3 Quantum measurement In a Stern-Gerlach experiment, for simplicity on a single silver atom with its spin perpendicular to the ®eld, SchroÈdinger's equation can be applied to the combined system of atom and photographic plate. The result is a quantum state that is a coherent linear combination of the quantum states of the two possible latent images of the atom on the plate. What is seen is quite different: an image in one place or the other. This is a simple example of the paradox of SchroÈdinger's cat [133, 132, 154]. It is an example of destructive measurement. In the case of nondestructive measurement, SchroÈdinger's equation produces an entanglement between the measured system and the measurer, with non-zero amplitudes for all the possible values of the relevant classical variables of the measurer. Again this is not seen. The classical variable has a more-or-less de®nite value, the entanglement for an individual run of an experiment is destroyed, but reappears as a correlation between measurer and measured system in an ensemble of many runs. A quantum measurement is an interaction between a quantum system and its environment in which the state of the quantum system signi®cantly in¯uences a classical dynamical variable of the environment. A laboratory measurement is just one example, but there are many others, described in section 3.5. The interaction is not instantaneous, but takes a ®nite time. Because of the interaction, the system is open, and we can apply QSD. The interaction is always complicated, but for the measurement of a single dynamical variable with operator F, the main effect of the interaction on the system can be represented by a single Lindblad operator L ˆ cF, where c is a positive constant that determines the rate of the measurement. At ®rst we suppose that the effect of the Hamiltonian evolution of the system can be neglected during the course of the measurement, as in a well-prepared Stern-Gerlach experiment. The system is then wide open. Because L is an Hermitian operator, the QSD equation is jd i ˆ ÿ12 L2
j idt ‡ L
j id;

…5:6†

where L
is the shifted operator de®ned by (3.22). For any Hermitian operator G that commutes with L, the change in the expectation hGi of G due to the diffusion of j i in time dt is

5.3 Quantum measurement

73

dhGi ˆ h jGjd i ‡ hd jGj i ‡ hd jGjd i ˆ 2Reh jGjd i ‡ hd jGjd i ˆ ÿhGL2
idt ‡ hL
GL
idt ‡ 2Re…hGL
id† ˆ 2Re……G; L†d†

…5:7†

ˆ 2…G; L†dR ; since …G; L† is real. We always use 2Re to represent the addition of the complex conjugate or the Hermitian conjugate to the expression that follows it, sometimes denoted ‡cc or ‡HC on the right of the expression in the quantum optics literature. For L ˆ cG;

we have

dhGi ˆ 2c  2 …G†dR :

…5:8†

Taking the mean, we get zero, M dhGi ˆ 0. The ensemble mean of the change in expectation is zero, so that the measurement of a dynamical variable does not change the mean of any dynamical variable that commutes with it. In particular it does not change its own mean, which is just as well, as the result would be wrong if it did. Equation (5.7) is used to obtain the changes in hGi and hG2 i that are needed for the change of the localization …G†, which is the inverse of the ensemble mean of the variance of G. The change in M  2 …G† is ÿ
M d 2 …G† ˆ M dhG2 i ÿ 2hGidhGi ÿ …dhGi†2 ˆ ÿ2‰…G; L†Š2 dt:

…5:9†

When L ˆ cG, this simpli®es to M …d 2 …G†=dt† ˆ ÿ2c2 ‰ 2 …G†Š2 :

…5:10†

To go further, look at the whole ensemble over a ®nite time. This requires an ensemble mean of both sides, giving d M  2 …G† ˆ ÿ2c2 M … 2 …G††2  ÿ2c2 …M  2 …G††2 ; dt

…5:11†

where the inequality follows from the positivity of the ensemble variance of  2 …G† de®ned by equation (2.8):

74

Localization

ÿ
2 ÿ
2 2 … 2 …G†† ˆ M  2 …G† ÿ M  2 …G†  0:

…5:12†

So a bound on the change of the ensemble localization of G ÿ1  …G† ˆ M  2 …G†

…5:13†

is given by d ÿ1 … †  ÿ2c2 ÿ2 ; dt

so that

d  2c2 dt

and

…t†  …0† ‡ 2c2 t:

…5:14†

In practice, the value of c is very large when the system becomes entangled with macroscopic variables, as it does during a measurement. Localization then appears to be instantaneous. This is the quantum jump limit of a measurement, giving a physical picture which is consistent with the Copenhagen interpretation. The above treatment of measurement makes no reference to the physics of the measuring apparatus. It is formal, and one of the aims of QSD is to show that such formal treatments follow from a more detailed theory of the interaction between the measured quantum system and the measuring apparatus as the environment of the system, as discussed in section 4.8. The theory depends on where the division between system and environment is made, and on the detailed physics of the measuring apparatus, including dissipation in ampli®ers and resistors. Section 6.5 contains a brief description of such a theory, as applied to the continuous Stern-Gerlach effect by Alber and Steimle. Quantum measurement is an untypical interaction between system and environment, which produces localization in only one dynamical variable of the system. For most systems, phase space localization is the norm. This has been demonstrated extensively by numerical experiments, and is analysed in the next section and in chapters 9 and 10. 5.4 Dissipation The simplest types of dissipation are represented by the rÿ spin operator and the annihilation operator a, which is used to illustrate phase space localization here. These are non-Hermitian operators, and they do not commute with

5.4 Dissipation

75

their conjugates. The theory is not quite so simple as for measurement, and we take the opportunity to derive the expression for the change in the variance  2 …G† of an arbitrary operator G for a wide open system with any number of arbitrary Lj . From the QSD equation (4.21) with H ˆ 0, after much cancellation, the change in hFi for arbitrary F and many Lindblads is dhFi ˆ h jFjd i ‡ hd jFj i ‡ hd jFjd i X y fhLj FLj ÿ 12Lyj Lj F ÿ 12FLyj Lj idt ˆ j

ˆ

X j

…5:15†

‡…F; Lj †dj ‡ …Lyj ; F†dj g f 12 hLyj ‰F; Lj Š ‡ ‰Lj ;y FŠLj idt ‡ …F; Lj †dj ‡ …Lyj ; F†dj g:

Putting F ˆ G and F ˆ Gy G we ®nd that M d 2 …G† ˆ M dhGy Gi ÿ 2M RehGy idhGi ÿ M dhGy idhGi ˆ 12 M hLyj ‰Gy G; Lj Š ‡ ‰Lyj ; Gy GŠLj idt ÿ M Re…hFy ihLyj ‰G; Lj Š ‡ ‰Lyj ; GŠLj i†dt ÿ M …j…G; Lj †j2 ‡

j…Lyj ; G†j2 †dt

(commutation) (commutation) (covariance): …5:16†

The commutation terms can have either sign, but the covariance term must always be negative. So a Lindblad L localizes an operator G for which the commutators vanish, unless the covariances are both zero. But if the operators do not commute, then the commutators can produce delocalization. This is what happens for conjugate variables, so that they satisfy Heisenberg indeterminacy despite the localization. However, in the classical theory of chapters 9 and 10, the commutation terms are negligible, and even conjugate variables are localized. Now apply the general result (5.16) to localization due to dissipation, when there is a single Lindblad proportional to the annihilation operator: 1

L ˆ 2 a;

G ˆ a;

…5:17†

where is a rate. The commutation relation ‰a; ay Š ˆ 1 gives h i ÿ
2 M d 2 …a† ˆ ÿ M  2 …a† ‡  2 …a† ‡j…ay ; a†j2 dt;

…5:18†

76

Localization

which is always negative unless both  2 …a† and …ay ; a† are zero. Using the inequality (5.12), dropping the last term in (5.18) above, and putting  ˆ 1=M  2 …a† gives d…1=†  ÿ …1= ‡ 1=2 †dt

or

d  … ‡ 1†:

…5:19†

To get the boundary of the inequality over a ®nite time, we replace it by an equality and integrate, to get …t†  ……0† ‡ 1†e t ÿ 1:

…5:20†

The localization increases exponentially at a rate given by , which is also the rate of dissipation. In the mean, the quantum wave localizes exponentially at this rate to a coherent state, which is a minimum indeterminacy wave packet. 5.5 Channels and statistical properties Channels help us to understand the physics of measurement, and to analyse the statistical properties of quantum states. The space of states j i of a quantum system may be divided or partitioned into orthogonal subspaces or channels. The channels are labelled by the complete set of projectors Pk that project onto them, where Pk P` ˆ P` Pk ˆ k` P` ;

X

Pk ˆ I:

…5:21†

k

The partitions are often spatially separated. In that case the projectors project onto mutually exclusive regions of position space, and the channels can be thought of as regions of ordinary position space. For the Stern-Gerlach experiment, position space is divided into two parts, each containing a beam of atoms with a ‡ or ÿ spin. More generally, channels correspond to the eigenspaces of Hermitian operators. When these operators represent dynamical variables, then the channels label values or ranges of those variables. A channel may contain one state, an in®nite continuum of states, or anything in between. We can ignore transitions between the internal states of each channel, and consider only the transitions between the channels, which are represented by relatively simple operators. A complicated interaction that produces localization in eigenspaces of a dynamical variable with Hermitian operator G can be represented by an Hermitian Lindblad cG. If the operators are chosen well, then the compli-

5.6 Localization theorems

77

cated physics of more general interactions can be represented adequately, though approximately, by simple models with a few Lindblads, not necessarily Hermitian. One simpli®cation is that each channel, consisting of any number of states, can be represented by a one-state channel. The transitions between channels are then represented by relatively simple operators. For example annihilation and downward transition operators can represent dissipation by the environment, and a complicated interaction that produces localization in eigenspaces of a dynamical variable can be represented by an Hermitian environment operator cG. If the operators are chosen well, then localization of quite complicated systems can be represented adequately, though approximately, by these simple models. 5.6 Localization theorems This section states two theorems about localization in different channels: the dispersion entropy theorem and the general localization theorem. The latter is proved here, and the former in the next section. Other properties of localization are derived in chapters 9 and 10 on classical and semiclassical theory. Suppose there are two or more channels de®ned by a complete set of independent projectors Pk . A block diagonal operator has no nonzero off-diagonal matrix elements coupling the channels, so the nonzero elements lie in square blocks along the diagonal, each block representing a channel, as in ®gure 5.4. For the

Figure 5.4 The matrix of a block diagonal operator. Unshaded areas are zero.

78

Localization

Figure 5.5 The matrix of a local operator. Unshaded areas are zero.

dispersion entropy theorem, we assume that the Hamiltonian interaction between channels is negligible compared with the localization rate. Consequently the Hamiltonian is block diagonal and for every channel projector Pk , we have ‰H; Pk Š ˆ 0: The local operator shown in ®gure 5.5 consists of only one diagonal block, with nonzero elements in one channel only. Because of principle Pri1 of section 5.1, the ¯uctuations in different regions of position space are independent. So for this case the Lindblads are local. This is the other condition for the theorem, from which it follows that every Lindblad belongs to a channel with projector Pk . Let the Lindblads belonging to a particular Pk be labelled Lkj with j ˆ 1; 2;. Since they are local, each one of them satis®es the condition P` Lkj P`0 ˆ `k `0 k Lkj

(all `; `0 †:

…5:22†

Neither the Hamiltonian H nor the Lindblads couple the channels. So the channels are separable, since each channel and its environment operate independently. Because of the separability, the ensemble mean MhPk i is conserved. Later we shall see that, for the individual systems of the ensemble,

5.6 Localization theorems

79

the weight Wk , which is the conditional probability hPk i for being in channel k, is not conserved. This is the nonlocal property of the localization process: despite the lack of interaction between the channels, which are often separated in space, the weights change. A normalized state vector j i can be represented as a linear combination of normalized channel state vectors j iˆ

X

ak j

ki

…h

kj ki

ˆ 1†;

…5:23†

k

where jak j2 ˆ hPk i ˆ Wk

…5:24†

is the weight of the state j i in the channel Pk . The expectations of the Lindblads are hLkj i ˆ h jLkj j i ˆ h

k jLkj j k i;

…5:25†

where Lkj is the jth Lindblad belonging to channel Pk . The effective interaction rate in channel k is then de®ned as Rk ˆ

X j

jhLkj ij2 ;

…5:26†

which has the dimensions of inverse time. The dispersion or delocalization for state j i and partition Pk is measured by the quantum dispersion entropy Q: Q ˆ Q…j i; Pk † ˆ ÿ

X k

Wk ln Wk :

…5:27†

Evidently, the smaller the entropy, the greater the localization among the channels. The dispersion entropy theorem states that the mean rate of decrease of quantum dispersion entropy for the partition is minus a weighted sum over effective interaction rates:

80

Localization

X 1 ÿ Wk d …M Q† ˆ ÿ Rk Wk dt k X ˆÿ Wk …1 ÿ Wk †…Rk =Wk2 †  0;

…5:28†

k

where Rk =Wk2 are nonsingular. For only two complementary channels,   d W2 R1 W1 R2 …M Q† ˆ ÿ ‡ W1 W2 dt ˆ

ÿW1 W2 …R1 =W12

‡

R2 =W22 †

…5:29†  0:

The mean quantum dispersion entropy always decreases, so the localization always increases, unless the effective interaction rate is zero, or the localization is complete. Section 5.8 discusses this result. The proof is given in section 5.7. The second theorem of this section is the general localization theorem, for which the conditions are weaker. The Hamiltonian can couple the channels and the Lindblads Lj are block diagonal in the channel subspaces, so that, for all Pk and Lj , Pk Lj ˆ Lj Pk ˆ Pk Lj Pk :

…5:30†

The general localization theorem for an initial pure state then says that for each projector Pk the change in the ensemble mean of the variance is dM  2 …Pk † ˆ M …1 ÿ 2hPk i†hÿi hÿÿ1 ‰H; Pk Šidt ÿ M …j…Pk ; Lj †j2 ‡ j…Lyj ; Pk †j2 †dt:

…5:31†

The Hamiltonian term follows from (3.2) and P2k ˆ Pk . The remaining terms follow from equation (5.16) with G ˆ Pk , where all sums but the last are zero because Pk commutes with Lj . If the Hamiltonian, like the Lindblads, is block diagonal in the subspaces of Pk , then the ®rst term is zero, and the system localizes in the space of each Pk : d M  2 …Pk †  0; dt

…5:32†

with equality when hPk Lj i ÿ hPk ihLj i ˆ …Pk ; Lj † ˆ 0

( for all k; j†;

…5:33†

5.7 Proof of the dispersion entropy theorem

81

which is a very special case. When the Hamiltonian is block diagonal, the expectation of each projector tends to zero or one, and the system can only localize into one of the channels. When a de Broglie wave enters a medium such as a solid, liquid or a dense gas, the Hamiltonian then represents the diffusion of the particle within the medium. The change in position as a result of the diffusion is on a microscopic scale during the localization, so even when the Hamiltonian is not zero, its effects are negligible, and a particle which interacts with material in a number of regions of space localizes into just one of them. An example is the localization of a silver atom by the photographic emulsion in the SternGerlach experiment.

5.7 Proof of the dispersion entropy theorem For the change in the dispersion entropy (5.27) we need to evaluate d…x ln x† to second order in dx, which is d…x ln x† ˆ …1 ‡ ln x†dx ‡ …dx†2 =…2x†:

…5:34†

Since M dWk ˆ 0, only the quadratic term contributes to the change in the entropy M d…Wk ln Wk † ˆ M …dWk †2 =…2Wk †

…5:35†

and so M dQ ˆ ÿ

X

M …dWk †2 =…2Wk †:

…5:36†

k

Now use equation (5.15) for the change in hPk i. Since the Lindblads are local, they are also block diagonal, and the commutators are zero. The quantum covariances are 

…Pk ; Lkj † ˆ …1 ÿ Wk †hLkj i ˆ …Lykj ; Pk † ; …Pk ; Lk0 j † ˆ ÿWk hLk0 j i ˆ …Lyk0 j ; Pk † so

…k0 6ˆ k†;

…5:37†

82

Localization

…dWk †2 ˆ …dhPk i†2 ˆ 2…1 ÿ Wk †2

X j

jhLkj ij2 dt ‡ 2Wk2

X 2

ˆ 2…1 ÿ 2Wk †Rk dt ‡ 2Wk

k0

XX k0 6ˆk

j

jhLk0 j ij2 dt

Rk0 dt: …5:38†

Therefore, since

P k

Wk ˆ 1,

X X X X 1 ÿ 2Wk d Rk ÿ M Wk R k0 : ÿ jhLk0 j ij2 M Q ˆ ÿM W dt k k k k0 j;k0 X 1 ÿ Wk 4ÿ Rk ; …5:39† Wk k which proves the dispersion entropy theorem. 5.8 Discussion First consider the four principles of section 5.1. Independent parts of the environment of a system usually act independently, and so are represented by different sets of environment Lindblads Lj , and statistically independent ¯uctuations dj . This is the principle Pri1, which can be applied to the diffusion of a particle by a gas, or the absorption of a particle by a screen. Localization needs interaction between system and environment, producing entanglement between them. One-way action of system on environment is not enough (section 4.6). Consequently the scattering of a particle by a ®xed scattering centre has zero effective interaction rate, since there is zero recoil, and if the scatterer is very heavy relative to the system, the effective interaction rate and resultant entanglement are very small. These are examples of action of the environment on the system, and not of interaction between environment and system. They can be represented exactly or approximately by an additional ¯uctuating Hamiltonian, not by Lindblad environment operators, and they do not produce localization. The same goes for the focussing and diffraction of photons by nonabsorptive optical apparatus, or the electrons in an electron microscope, where the recoil of the apparatus is negligible. The principle Pri2 follows directly from the state diffusion equation (4.21). Amplitude ¯uctuations always occur unless j i is a simultaneous eigenstate of all the Lindblads, which happens when

5.8 Discussion

Lj j i ˆ hLj ij i

83

(all j†

…5:40†

and which is clearly exceptional. If the Hamiltonian H and the Lindblads Lj are block diagonal, so that they do not couple the channels signi®cantly, the ensemble probability of being in any of the channels is conserved. But for the state j i of an individual system of the ensemble, the magnitude and phase of the channel components Pk j i all ¯uctuate, unless (5.40) is satis®ed. There is more detail in the rest of this chapter. Pri3 is a consequence of the localization theorems. Pri4 is about the special case of position localization. In that case the projectors Pk project onto mutually exclusive regions of position space, and, provided that these regions are not too small, the Hamiltonian coupling between the different parts of the environment in the different regions can be neglected. Further, it is a consequence of the approximate locality of interactions that the environment of the region Pk has no effect on the system when the system is in a different region Pk0 , and so the conditions of the dispersion entropy theorem are easily and very generally satis®ed for the position variable. The permanence of position localization depends on the standard of rest, or rest frame. In any frame that moves signi®cantly with respect to the physical environment, localization is not permanent, because the system will necessarily move from one region to another, spoiling the position localization. So the environment provides the reference frame for position localization. We have shown how the quantum state diffusion equations of an open system represent explicitly the process of localization in a subspace or channel of the state space. The localization due to the environment increases when the ensemble mean of the quantum variance or the quantum dispersion entropy of the channel projectors decreases, and the two theorems demonstrate that this always happens unless the change is zero. Explicit formulae for the rates of change are obtained in terms of effective interaction rates. Localization is characteristic of the interaction of a system with its environment, whether or not that environment contains measuring apparatus or an observer. It may seem remarkable that the mean quantum dispersion entropy should behave contrarily to every other physical entropy under the conditions given here. But this is just a consequence of the remarkable properties of quantum mechanics, in which wave properties of a system, which require it to be extended, are followed by particle properties, in which it is localized, as for

84

Localization

example in the two-slits experiment. The dispersion entropy theorem is a mathematical expression of these properties. The reduction of mean dispersion entropy is consistent with the increase of thermodynamic entropy because ensemble entropies increase as a result of the diffusion of pure quantum states, and this more than compensates for the decrease in the mean over the quantum dispersion entropies of the individual pure states of the ensemble. The Hamiltonian term can increase the mean quantum dispersion entropy, and often does, so in general this entropy can either increase or decrease. This contrasts with the entropy theorem of the classical theory in chapter 9. In the QSD picture the localization process appears explicitly, whereas in the usual picture the density operator gives an average over the localized states. The reduction of its off-diagonal elements due to the environment has been associated with the localization process before, as described in detail by Zeh and his collaborators, and also by OmneÁs. This view is further discussed in chapter 7. Depending on the nature of the system and its interaction with the environment, there may be localization in many different dynamical variables, but localization in position is particularly important because interactions are localized in position space, but not in momentum or other dynamical variables. For compound systems with many particles, the interaction is not localized in the con®guration space of the system, but the interaction of each distinguishable particle with the environment is localized in position space, and the overall effect is to localize each particle in position space, which is equivalent to localizing the whole system in con®guration space.

6 Numerical methods and examples 6.1 6.2 6.3 6.4 6.5 6.6 6.7

Methods Localization and the moving basis Dissipative quantum chaos Second-harmonic generation Continuous Stern-Gerlach Noise in quantum computers How to write a QSD program

After a brief summary of the methods that can be used to study open quantum systems, the formulation and numerical methods of QSD are described, with particular emphasis on the moving basis. This is followed by summaries of various applications which illustrate the use of the program and show something of the variety of problems that can be tackled using QSD. The ®nal section explains how to obtain and use a general-purpose QSD computer program, using the computations needed to prepare some of the ®gures in chapter 4 as examples. Key references are [127, 130]. 6.1 Methods The problems are to simulate the evolution of an open quantum system, and to compute the properties of an ensemble of open quantum systems. Numerical solutions are needed for all but the simplest systems, because analytic solutions cannot be found. All the methods described here depend for their formulation on the Lindblads. Fortunately, past research on master equations has produced a large stock of Lindblads for a variety of processes and systems, which can now be used for QSD. The same past experience is very valuable for the choice of boundary between system and environment, described in section 3.6. There are several numerical methods (NM) for solving these problems, including quantum state diffusion: NM1. Master equation. Direct stepwise integration for the ensemble. This is expensive in computer resources, but provides the most accurate results when it is possible to use it. It does not simulate the evolution of individual systems.

85

86

Numerical methods and examples

The remaining methods are based on the unravelling of the master equation in different ways into an ensemble of pure states. They are sometimes known as quantum trajectory or Monte Carlo simulation methods. NM2. Quantum jump methods. These involve continuous integration of the drift term, including the SchroÈdinger term, interrupted by discrete quantum jumps representing the stochastic part of the interaction with the environment, usually considered as a measurement process. These have been the most widely used in quantum optics, where they were developed. They are at their best when the main environmental effect is a measurement involving particle counting, normally photon counting, or a process which is easy to simulate in this way. NM3. Quantum state diffusion, QSD. This is often preferable when the environment changes continuously instead of by discrete jumps, and when its in¯uence on the system is well represented by a known master equation. QSD is particularly ef®cient when there is strong localization. NM4. The moving basis. This is also known as mixed classical-quantum representation. It is not really a different method, but a supplement to the quantum trajectory methods, in which the basis states are made to change with time. It makes a big improvement in the ef®ciency of these methods when there is signi®cant localization in phase space and is very effective as the motion approaches the classical limit. It is of little help for the master equation.

Quantum jump methods now have an extensive literature, some of it referenced in the introduction and in section 4.7, so this chapter concentrates on QSD, using an easily available computer library prepared by Schack and Brun [130], which we will call the QSD library. The QSD library can also be used for jump methods and for mixtures of diffusion and jumps, where appropriate. The solution of a time-dependent SchroÈdinger equation for an isolated system, or a QSD equation for a single open system of an ensemble, is the trajectory of a pure quantum state in Hilbert space. The computer representation of the state j …t†i at a given time t requires a ®nite basis of (usually) orthonormal states jun i such that j iˆ ˆ

nx X nˆ1 nx X nˆ1

jun ihun j i nx X cn jun i ˆ …cnR ‡ icnI †jun i:

…6:1†

nˆ1

The number of basis states nx may be the dimension of the state space, for example when the system consists of a ®nite number of spins. But usually it is

6.1 Methods

87

chosen as an approximation to a state space of in®nite dimension, for example when the Hamiltonian and Lindblads include positions, momenta, or ®eld amplitudes with in®nite spectra. The representation of the state is then approximate too. The state is represented in the computer by the 2nx real numbers cnR ; cnI . The evolution of the system over a time t is simulated by matrix multiplications of state vectors, where the matrices represent the Hamiltonian and Lindblad operators. In general they are nx  nx matrices, but whenever possible the representation is chosen so that the matrices are sparse, so that the number of non-zero or signi®cant elements is proportional to nx . This reduces the time of computation by a factor of order nx , but puts a severe constraint on the choice of representation. Computations are limited both in space (computer store or memory) and in time. Despite enormous advances in computer technology, these limitations are still important. For a single trajectory, only a few operations are performed on each number cn …t† for a given time t. The time of computation for a single quantum trajectory depends roughly linearly on the space used in the computer store, and so we only need consider the space. Even when the Hamiltonian and Lindblad operators are represented by sparse matrices, the density operator rarely is. The store for a state vector needs to hold about 2nx real numbers, but the store for a density operator needs to hold n2x real numbers. So the solution of the master equation almost always needs much more space and time than a single run for a trajectory. In practice we ®nd that even a few runs of the QSD equations can tell us a lot about an open system, just like a few runs of a laboratory experiment. This analogy still holds when the system is far away from any laboratory, as for an atom in the atmosphere of a star. In fact, running QSD on a computer is very similar to running a laboratory experiment. In such situations QSD equations gain over master equations by a factor of order nx in computer space and in time. In practice it is usually found that the space limitation is more severe. However, for many purposes, ensemble means are the signi®cant quantities, in which case QSD requires many runs with different ¯uctuations in order to build up statistics. In this mode QSD is a Monte Carlo method, with all its advantages and disadvantages. It is like a computer model of a laboratory experiment in which the number of runs is suf®cient to build up `good statistics'. A computer experiment, like a laboratory experiment, needs careful design and planning, and the runs take a lot of time. Under these circumstances numerical solution of the master equation is always more accurate and sometimes quicker, provided it is possible to use it. But very

88

Numerical methods and examples

often it is not possible, because there is not enough space in the computer. Simulation methods like QSD or quantum jumps are then the only ones available. 6.2 Localization and the moving basis For systems with ®eld modes or moving particles, localization often gives us a method of further improving the ef®ciency of computation by a large factor. Suppose the system has operators with corresponding dynamical variables of d freedoms and a 2d-dimensional phase space with vector coordinate and momentum …q; p†. The volume of a region of phase space has dimensions of (action)d , which can be measured in Planck units of …2 hÿ†d , the volume of a phase space Planck cell. This measures the number of well-chosen basis states needed to represent a state vector which is con®ned within the region [109]. We need a measure of the approximate effective volume V of phase space occupied by a pure state j i. This is provided by taking the overlap with standard shifted wave packets, coherent states uq0 p0 , for which the expectations of their d-dimensional coordinate and momentum are …q0 ; p0 †. The effective volume V is de®ned as the volume for which the overlap is greater than some prescribed small value 0 , as illustrated in ®gure 6.1. The circles of area 2 hÿ in 2 the ®gure represent coherent states ji with overlap jh jij2 ˆ 0 . See section 10.5 for more about coherent states and their representation. For typical Hamiltonians and Lindblads, an approximate position representation, or alternatively the number states of a harmonic oscillator, give suitable bases, in which the operators representing dynamical variables are

Figure 6.1 The effective volume V in phase space.

6.3 Dissipative quantum chaos

89

represented by sparse matrices. Since most applications have been in quantum optics, where the ®eld modes are oscillators, and nonlinearities are usually small, the oscillator basis is used here and in the computer library of section 6.7. The minimum number of states required to represent j i to a given precision is then given approximately by the number of Planck cells in V. As the state of a system localizes in phase space, fewer basis states are needed. If the localization is suf®ciently strong, V may be reduced to a few Planck volumes, with a corresponding small number of basis states. This small volume moves around in the phase space along an approximately classical trajectory, so, to take advantage of the localization, the basis must follow the trajectory. The detailed theory is given in chapter 10. This is the moving basis [146, 127]. The volume of phase space occupied by the ensemble is almost always much bigger than the volume occupied by the pure states, and so localization rarely works for the master equation. This is shown particularly clearly in the example of the chaotic motion of the Duf®ng oscillator in the next section. In practice, the current values of the expectations hqi; hpi, which de®ne the centroid of the wave packet, are used as the centre of an oscillator basis, with Fock states jni obtained by shifting from a standard oscillator basis centred at the conventional origin of the phase space. States are included up to some maximum value nx . With an appropriate choice of conjugate coordinates and momenta, these cover a circular or 2d-dimensional hyperspherical volume of phase space up to some precision which is determined by the program. To maintain precision, the hypersphere must include the volume V, as illustrated in ®gure 6.2. 6.3 Dissipative quantum chaos Now for speci®c applications, using QSD with a moving basis. Classical chaos in dissipative and Hamiltonian systems has been widely studied by numerical integration of the equations of motion for a sample of trajectories. The quantum chaos of Hamiltonian systems has also been well studied, mainly by analysing spectra. The nature of the classical limit for Hamiltonian systems is a subtle one. By comparison, quantum dissipative systems have been neglected. This is partly because the density operator does not represent individual systems. Quantum trajectories represent the evolution of individual systems. They are the quantum analogue of classical trajectories, and localization makes the classical limit of QSD relatively straightforward in principle. In practice it is a numerical challenge, because, as the ratios of the classical actions to Planck's

90

Numerical methods and examples

Figure 6.2 Shifted oscillator eigenstates and the volume V:

constant increase, so does the effective size of any ®xed basis used to represent the quantum states. The moving basis is ideal for this situation. It becomes relatively more and more ef®cient as the classical limit is approached. Spiller and Ralph [140] were the ®rst to study the emergence of chaos in open quantum systems, in a dissipative quantum map. Here we discuss the work of Brun and his collaborators [21, 77], who used the QSD program library of section 6.7 to look at the classical limit of the quantum Duf®ng oscillator. The classical Duf®ng oscillator is a damped, driven, one-freedom system with the double well potential x4 x2 V…x† ˆ ÿ ; 4 2

…6:2†

a driving force of the form g cos…t† and a damping ÿ2ÿx_ proportional to the velocity. The classical equations of motion are d2 x dx ‡ 2ÿ ‡ x3 ÿ x ˆ g cos…t†: 2 dt dt

…6:3†

Because of the time dependence, the solutions lie in a phase space of three dimensions, and the evolution is conveniently represented by the intersection of the trajectories with a Poincare surface of section. It has a chaotic regime with a strange attractor.

6.3 Dissipative quantum chaos

91

For the master equation and for QSD, the Hamiltonian is Hˆ

p p2 2 q4 q2 g ‡ ÿ ‡ cos…t† q ‡ ÿ…qp ‡ pq† 2 4 2

…6:4†

and there is one Lindblad, which is p p L ˆ 2 ÿa ˆ 2ÿ…q ‡ ip†;

…6:5†

where a is the annihilation operator. We put hÿ ˆ 1. The constant is a scaling factor, which increases the scale of the classical actions as ! 0, whilst keeping hÿ constant, so this corresponds to the classical limit. The intersections of the QSD phase space trajectory hq…t†i; hp…t†i with the surface of section are illustrated in ®gure 6.3, for ÿ ˆ 0:125; g ˆ 0:3 and with four different values of . The quantum system with ˆ 1 can be represented by 10 basis states. So it can easily be solved using any method, including the master equation. The chaotic behaviour is entirely masked by quantum ¯uctuations of the centroid. QSD with a moving basis was used to investigate the transition from this quantum regime to the classical limit, which is a far more demanding computational problem. For ˆ 0:01 the quantum surface of section is indistinguishable in form from the classical plot, and the intermediate values of clearly show the transition from quantum to classical dynamics. The ®nal effective value for hÿ for (d) was scaled down by a factor of 10 000 compared with its initial value for (a). The density operator would have needed more than 109 real numbers, and a ®xed basis about 50 000 real numbers. For this near approach to the classical limit, QSD with a moving basis needed about 15 basis states. The same system has been studied in the adiabatic or rotating wave approximation by Rigo, Alber, Mota-Furtado and O'Mahony [124]. They use QSD to show that an individual quantum system can exhibit bistability, going over smoothly to the classical limit, with a mean transition time for the switching between the two equilibrium points due to quantum ¯uctuations. This resolves a controversy concerning this limit.

Figure 6.3 The Poincare surface of section for a single QSD trajectory of a forced damped Duf®ng oscillator in the chaotic regime with ÿ ˆ 0:125, g ˆ 0:3 and four scalings: (a) ˆ 1:0, (b) ˆ 0:25, (c) ˆ 0:1 and (d) ˆ 0:01:

6.4 Second-harmonic generation

93

6.4 Second-harmonic generation For d freedoms and N basis states per freedom, the total number of basis states required is of order N d , so two freedoms are much more of a challenge than one. Second-harmonic generation or frequency doubling is a standard process of quantum optics which produces squeezed light (see section 10.5). It is a system of two freedoms, which are optical modes or oscillators, with angular frequencies !1 ; !2 , approximately in 2 to 1 resonance. They interact through a nonlinear coupling in a cavity driven by a coherent classical external ®eld of angular frequency !f , which is in near resonance with !1 . The cavity modes are slightly damped and detuned, with detuning parameters 1 ˆ !1 ÿ !f ;

2 ˆ !2 ÿ 2!f :

…6:6†

Thus the cavity drives the ®rst oscillator, which then drives the second through the coupling. From the the interaction representation master equation formulation of Drummond and his collaborators [43, 44], the Hamiltonian is given by H= hÿ ˆ 1 ay1 a1 ‡ 2 ay2 a2

(oscillators)

‡ i…ay1 ÿ a1 †

(driving force)

2 y ‡ 12…ay2 1 a2 ÿ a1 a2 †

(coupling);

) …6:7†

where the aj are annihilation operators. The damping is represented by the Lindblad operators q Lj ˆ 2j aj :

…6:8†

A direct numerical solution of the master equation for this problem is dif®cult, because in the usual Fock basis the number of basis states needed is equal to the product of the highest signi®cant photon numbers nj in each mode, and this can become very large. The QSD method with a ®xed Fock basis was used by Gisin and Percival [72] and by Goetsch and Graham [78]. Zheng and Savage applied the quantum jump method without a moving basis to the chaotic regime of this system with typical photon numbers (oscillator levels) of 200 in each mode [159]. About 500 basis states per mode were used: roughly 250 000 basis states in all. This was a tour-de-force which required hundreds of hours on a 32-processor supercomputer.

94

Numerical methods and examples

Figure 6.4 The expectation of the photon number versus dimensionless scaled time for a single second-harmonic generation trajectory. Full curve: fundamental mode hn1 i. Broken curve: second harmonic mode hn2 i. The values of the parameters are given in [128].

The same problem was studied using QSD with a moving basis by Schack and his collaborators [127], where more references to earlier treatments may be found. However, the mean photon number was chosen to be about six times larger in each mode, making it even more dif®cult to solve. The total number of basis states depended on the precision required, but was typically about 200. It took a few hours per trajectory on a PC, and the calculated variation in mean photon number of each oscillator as a function of time is shown in ®gure 6.4. The big reduction in the number of basis states was due to the use of the moving basis. In [128], QSD was used on the same problem to compute optical spectra. 6.5 Continuous Stern-Gerlach In one of the most detailed QSD computations of a laboratory measurement, Steimle and Alber have analysed the continuous Stern-Gerlach effect using QSD with a moving basis [145]. This study is particularly interesting because it makes a connection between the ®nite measurement times of QSD and the ®nite measurement times proposed by Dehmelt [27, 28] in his interpretation of experiments on continuous nondestructive measurements of the z-motion and spin of a single electron in a Penning trap. It takes account of environmental effects that are due to the measurement, and others that are not.

6.6 Noise in quantum computers

95

In contrast to the original Stern-Gerlach experiment described in section 5.1, the detection process does not absorb the electron, so its dynamics can be monitored continuously, and has been used in experiments to determine fundamental physical constants. The apparatus consists of a Penning trap containing a single electron which undergoes cyclotron, magnetron and zoscillations of very different frequencies, and an !z -shift spectrometer. This has a near-resonant sinusoidal drive which forces the z-motion of the electron into a state of high mean quantum number. This motion is coupled to an external circuit which has a resistor. The dissipation due to the resistor localizes the z-motion into an approximately coherent state whose phase is measured by a phase-sensitive detector. The cyclotron motion is coupled to the electron spin through weak inhomogeneous magnetic ®elds. The nondissipative dynamics is represented by a complicated Hamiltonian, which is simpli®ed by using the adiabatic approximation on the hierarchy of frequencies. The Lindblads represent the effect of the surrounding thermal radiation ®eld on the cyclotron motion, the dissipative effect of the resistor on the electronic z-motion and the continuous measurement of the out-of-phase quadrature component of the current induced in the resistor. Alber and Steimle's computations used parameters similar to those of the experiments. They showed that it was practicable to vary the effective measurement time, as de®ned by Dehmelt, over a wide range of values, corresponding to `complete' and `incomplete' measurements. They also showed numerically that, in the stationary limit, the ensemble mean of a quantum variance using the QSD wave functions remained essentially unchanged for two different choices of the boundary between system and environment. This example demonstrates by a tour-de-force calculation how questions of interpretation are related to the practical analysis of real laboratory measurements. 6.6 Noise in quantum computers Quantum computers are computers that depend on quantum mechanics for their operation [11, 12, 51, 52, 1, 31, 13, 138]. They are very promising [135], but they have not yet been built. The noise produced by interactions with the environment is a major problem. A program based on the QSD library has been used to simulate the effects of noise in quantum computers. Qubits are introduced in section 3.1. A quantum computer operates on qubits as a classical computer operates on classical bits. An individual qubit is represented in the computer by the state of a single two-state system, such

96

Numerical methods and examples

as a spin-half system. The binary numbers b10110 and b11011, representing the decimal numbers d22 and d27, are stored as the product states jd22i ˆ j1ij0ij1ij1ij0i

jd27i ˆ j1ij1ij0ij1ij1i:

…6:9†

The two numbers are stored together by the same ®ve spin systems as the linear combination p …1= 2†…jd22i ‡ jd27i†;

…6:10†

which is an entangled state of the qubits. More numbers can be stored in more elaborate entangled states up to the 32 dimensions of the Hilbert space of the ®ve spins. Numerical processes on these states are parallel operations on all the numbers. So quantum computers are strongly parallel computers that take full advantage of the dimensionality of the Hilbert space of quantum states. The Hilbert entropy, de®ned later in section 7.5, measures the information capacity of the binary elements of a quantum computer. Extracting information from a quantum computer is not so simple as extracting it from a classical computer, but it has been proved that if the noise problem can be solved, quantum computers would be far more powerful than classical computers for some problems, like factorization of integers into products of large primes [135]. Quantum error correction methods [136, 143, 144] can be used to reduce the problem, but the correction devices can themselves act as sources of noise. A multiple spin version of QSD was used by Barenco and his collaborators to support an analytic estimate of the effects of different kinds of noise, which showed that error correction can remain bene®cial in the presence of suf®ciently weak noise [5]. Their method of error correction for a single qubit over a given lapse time was this. First entangle it, using a coupling Hamiltonian, with four other qubits, then wait for a while with no Hamiltonian, and then disentangle the qubits using another coupling Hamiltonian to produce the wanted output qubit, so that the total time was equal to the lapse time. The noise present during the entire lapse time was represented by Lindblads operating on each of the qubits. The output state j 0 i was compared with the input state j i using the `mismatch' M ˆ 1 ÿ jh j 0 ij2 and the result used to check an analytic estimate, which was found to work well for reasonable values of the parameters. The dif®culty of simulating the quantum computer on a classical machine was itself evidence of the power of quantum computers. In this case it was possible to solve the master equation for 5 qubits, and the comparison with

6.7 How to write a QSD program

97

QSD showed that the latter only had statistical errors, and these were of the order of 1±2%. As the number nq of qubits increases, the computer space and time taken by QSD increases approximately linearly with nq , but for the master equation integration they increase at least quadratically, and soon swamp the space available. 6.7 How to write a QSD program The QSD program library written by Schack and Brun [130] is available on the World Wide Web. It can be obtained by downloading it from the site http://www.ma.rhbnc.ac.uk/applied/QSD.html

…6:11†

It is written in the C++ language, but for many purposes only the elements of C++ presented here are needed. The library was used to produce many of the ®gures of this book. The following description and the program 6.1 show in detail how that was done for ®gure 4.1a, as an example of how to write QSD programs using the QSD library. There is no single program. The details of the numerical methods are written in the library, and can be called when needed. For many purposes only a sketchy knowledge of C++ is needed to write programs using the library of main program templates. In C++ new types of variable can be de®ned, and the usual arithmetic operations like addition and multiplication can be de®ned for these new types, so programming more closely resembles the usual mathematical notation, and the use of the program library has some resemblance to the use of Maple or Mathematica. The library comes with a number of main programs which can be used as templates to write a wide variety of user programs. They are simple.cc, template.cc, onespin.cc, spins.cc, moving.cc, sums.cc.

…6:12†

The following shows how to use onespin.cc as a template to produce the graph of a Bloch sphere trajectory like that in ®gure 4.1a. Figure 4.1a illustrates the evolution of a two-state system, and the natural choice of template is onespin.cc, which solves QSD equations for such systems. The user programs are ®rst stored for safety in a separate directory. Then onespin.cc is retrieved into the main directory qsd1.3.2 (or later edition) for use as a template. This template is presented as program 6.1.

100

Numerical methods and examples

The ®nal edited form is shown as program 6.2. The output columns of the standard output, sent to a ®le (using standard output redirection) were used as coordinates for the graph of the trajectory on the Bloch sphere, without the sphere and lines of latitude and longitude. In ®gure 4.1a, the Hamiltonian was zero, so it is rede®ned as H = null; The number of Lindblads is 1, so it is kept as const int nOfLindblads = 1; The value of the coef®cient of the Lindblad is 0:135, so the variable gamma is rede®ned as double gamma = 0.135; To measure the z-spin, the Lindblad is Operator L1 = gamma*sz; When editing the main program, declare every new constant or variable. It must be one of the types already de®ned in one of the classes, for example a state, a basic operator or an operator. The main programs of the library are a good guide. Two constants, one real and one complex, are needed to de®ne the initial state: double TWOPI=2*3.14159; // required constant Complex I = (0.,1.); // required constant. The initial state is closer to the north pole than the south pole, and is given by State psi0(2,SPIN); // ground state (``spin down'') psi0 = cos(TWOPI/6)*psi0+(I*sin(TWOPI/6))*sp*psi0; // In N hemisphere, latitude 30deg // sp*psi0 is ``spin up''. Scalars multiplied first. The random number generator can be left as it is for one trajectory, but a new seed can be chosen for other trajectories. The basic time interval is suf®ciently small to illustrate a trajectory, but for a continuous line output

6.7 How to write a QSD program

101

it is advisable to have a shorter time interval between outputs. To illustrate localization, a longer total integration time is advisable. So this section of the program becomes // Stepsize and integration time double dt=0.01; // basic time step int numdts=2;

// time interval between outputs = numdts*dt

int numsteps=4000; // total integration time = numsteps*numdts*dt

…99†

The output variables, which give the coordinates on the projected Bloch sphere are sz,sx, and so the output becomes // Output const int nOfOut = 2; Operator outlist[nOfOut] = fsy,szg // Operators to output: This gives the standard output, which is used for the graph. For consistency, the output ®les should also be rede®ned as char *flist[nOfOut] = ``sy.out'',``sz.out''; // Output files There are two types of output. For each output operator G there is an output operator data ®le with ®ve columns. The ®rst column, labelled 0, is the time t. The rest are RehGi, ImhGi, Re‰hG2 i ÿ hGi2 Š, Im‰hG2 i ÿ hGi2 Š. The columns are treated as if labelled from 0 to 4. This is individual labelling. In all the output operator data ®les there are 1 ‡ 4nOfOut independent columns, labelled from 0 to 4nOfOut, with column 0 representing the time, and the remainder being the other columns of the output data ®les, taken in the order of appearance in the `outlist' of the program. This is standard output labelling, and is described in the comments which come with the program. The standard output contains a selection from these 1 ‡ 4nOfOut columns. The ®rst column labelled 0 is the time. The next four are a selection from the output operator data ®les, which are numbered according to the standard output labelling, as arguments in f; ; ; g in the `int pipe[4]' program line. This part of the standard output can be sent to a ®le for graphing. Column 6 is `dim', the current effective dimension of Hilbert space, that is the number of basis states used at the current time. Column 7 is `steps', the current

102

Numerical methods and examples

number of adaptive steps taken, which can be used to estimate the number of steps required by the program for a given physical time. For our standard output, the value in the third column, labelled 2 and representing hrz i is plotted against the value in the second column, labelled 1 and representing hry i. Further details of the program are given in [130]. More sophisticated changes, like introducing new classes, or programming the classical theory, cannot be made by editing the template main programs, and need a greater understanding of C++.

103

7 Quantum foundations 7.1 7.2 7.3 7.4 7.5 7.6 7.7

Introduction Matter waves are real Niels Bohr and Charles Darwin Quantum theory and physical reality Preparation of quantum states Too many alternative theories Gisin condition

This chapter and the next apply the methods of quantum state diffusion to the foundations of quantum mechanics. Problems with the foundations are ®rst introduced and then compared with similar problems in other ®elds. Quantum state diffusion has developed in parallel with its application to quantum foundations. The relation between them is presented through a selective history of alternative quantum theories that add stochastic terms to the SchroÈdinger equation. The reasons for introducing these theories, their scope and the constraints that apply to them are discussed. 7.1 Introduction The foundations of a ®eld of science are not like the foundations of a building, for the ®rst can survive reasonably well without proper foundations, but the second cannot. Nevertheless, the foundations of a ®eld of science can be important, and they are particularly important when the ®eld is being extended into completely new areas. This was true of quantum theory in the 1920s and 1930s and it is true of quantum theory today. The book by Wheeler and Zurek [154] contains an excellent collection of historical articles on the foundations of quantum mechanics. The quantum theory books by David Bohm [15], by Peres [119] and by Rae [123] use ordinary quantum theory and take particular care with the foundations. Much that is presented in this chapter follows Bell [10]. The quantum dynamics of atoms is very different from the classical dynamics of planets or the dynamics of the classical variables of laboratory apparatus. Seeing this difference in the 1920s, Niels Bohr and his colleagues in Copenhagen and GoÈttingen divided the universe into the classical and

104

Quantum foundations

quantum domains, with different physical laws in each. This was an essential feature of Bohr's Copenhagen interpretation of quantum theory. In Bohr's time, and for most physicists since then, the precise location of the boundary between these domains was of no importance. Atoms, molecules, baryons and leptons are quantum systems, whereas stars, planets, spectroscopes and accelerators are classical. The instruments are around 1030 times more massive than the systems they are used to study. The experimenters of the time had no possibility of locating the classical-quantum boundary, so its location was irrelevant in practice. The Copenhagen interpretation was a glorious and successful compromise, and a powerful theoretical tool for almost all quantum physics since. Classical and quantum systems obey different deterministic dynamics, but there is an essential indeterminacy where they interact. In 1935, those like Einstein and SchroÈdinger who were dissatis®ed with this division of the universe found themselves working at the margins of physics, because their objections made no contact with experiment or observation [47, 131]. Throughout most of the nineteenth century there was doubt about the reality of atoms. According to Ostwald and Mach in 1895, atomic theory was just a convenient model that could be used to answer questions about experiments in chemistry and physics, but those who believed in the reality of atoms were misguided [120]. By the end of the century most scientists had been convinced of their reality, but there were still many who disagreed. Chapter 2 of this book describes how Einstein and Perrin ®nally resolved the controversy with theory and experiments on Brownian motion. Atoms are real. Throughout most of the twentieth century there has been doubt about the reality of matter waves, particularly those in more than three dimensions. According to the quantum theory of many textbooks, they are just a convenient model that can be used to answer questions about experiments in physics and chemistry, but those who believe in the reality of these waves are misguided. This view is embedded in the language of those textbooks that tell us about quantum wave functions but not about quantum waves. But there are now many alternative theories in which all matter consists of quantum waves in three or more dimensions, with very different dynamical behaviour in the classical and quantum limits for large and very small systems. Since matter is real, so are the waves. In some of these theories, there is a diffusion, resembling a Brownian motion, of quantum states. There is today no experimental evidence like Brownian motion for the reality of matter waves, because we still have not been able to experiment at the

7.2 Matter waves are real

105

classical-quantum boundary. However, there is other experimental evidence, as we shall see. Physical reality was important for both Einstein and SchroÈdinger. Einstein, Podolsky and Rosen wrote the famous EPR paper on elements of physical reality in 1935 [47]. EPR also argued strongly for the locality of physical in¯uences. John Bell showed that locality and quantum theory were incompatible. Quantum nonlocality predicts distant correlations that are classically impossible, and so could be tested experimentally. Subsequent experiments, notably by Freedman and Clauser in 1972 [53] and by Aspect, Dalibard and Roger in 1982 [3], have provided evidence for quantum nonlocality in the laboratory. Details of these and related experiments are given in [119]. In these situations, physical systems behave nonlocally. So any alternative theory in which quantum systems like atoms and classical systems like measuring apparatus have the same dynamics has to be a nonlocal theory. SchroÈdinger was happy with his picture of an electron in a hydrogen atom as a wave in three-dimensional physical space, but not with the two electrons of a helium atom as a wave in a six-dimensional con®guration space. He could not accept the reality of the higher-dimensional waves. So in 1935 neither Einstein nor SchroÈdinger could develop an alternative realistic quantum theory of all matter because of their other unshakeable beliefs. Dirac wrote in 1963 that these problems with quantum foundations were real, but were not then ripe for solution, and so should be left for later [41]. Now modern instruments control individual atoms beyond Bohr's wildest dreams. New experiments are beginning to probe near possible boundaries between the quantum and the classical domains. New theories unite these domains, and some might be tested experimentally. So the problems of quantum foundations may now be ripe. 7.2 Matter waves are real Today's quantum nonlocality experiments make distant correlations that are classically impossible. Quantum cryptography keeps secrets in nonclassical ways. Section 6.6 explains how quantum computers, if they can be made, will also use higher-dimensional waves. For a long time we have known from Heisenberg's uncertainty principle that we can sometimes do less with a quantum system than expected from classical theory, but now we can sometimes do more. If higher-dimensional quantum waves were not real, how could we use them to do things that can't be done classically? Higher-dimensional matter waves must be real.

106

Quantum foundations

Modern physics demands universal theories. Since quantum waves are real and classical systems are real, we need a uni®ed dynamics for both. Without it, there is a gap in our physics, which cannot be ®lled till we understand the physics at the quantum-classical boundary and which will not be complete without experiments that explore the boundary. Atom traps and interferometers now provide experimenters with remarkable control over individual atoms. The classical-quantum boundary might be in their sights within a few years, so theory is needed to guide the experimenters. The `measurement hypothesis' of the usual theory is not enough. We need an alternative theory which describes the dynamics of classical and quantum systems as solutions of the same equations. The distinction between classical dynamics and quantum dynamics must come from limiting properties of the solutions of these equations and not from an evasive interpretation of the theory. SchroÈdinger's equation and its sophisticated ®eld and string theory extensions have to be modi®ed if they are to represent localized particles that obey classical dynamics. Several alternative quantum theories do this for the SchroÈdinger equation. For them localization is a dynamical process derived from a modi®ed equation. We need uni®ed theories of classical and quantum mechanics to guide the experimenters: where should they look for the elusive boundary between classical and quantum systems, and what should they look for? 7.3 Niels Bohr and Charles Darwin Probability, stochastic processes, and their problems are at least as important in biology as in physics. They are compared in [75]. In the theory of the evolution of species, stochastic processes come in universally as the source of new variation. They also represent the mixing of genes in sexual reproduction, but that does not concern us here. In Darwin's 1859 theory of evolution, the process of natural selection acts on whatever variability between individuals there is, to produce new varieties and species, but he was not clear on what caused the variability in the ®rst place. The mechanism became clearer with Mendel's 1865 rules of heredity, through what we now call genes, which gradually became accepted after 1900. The genetic theory was successfully incorporated into a general mathematical theory of evolution by natural selection in the 1940s. The new variation is produced by random mutation of genes, which has to be very small over a few generations. What are the genes? It was not until the DNA double helix containing the genes was discovered in 1953, that the physical basis of genetic variation could be understood.

7.4 Quantum theory and physical reality

107

Compare this with quantum theories of measurement. Although the mathematical laws of probability in the Copenhagen theory are clear, it is no more clear about the mechanism and the physical basis of the indeterminacy of quantum measurements than Darwin was about the mechanism and physical basis of variation in species. Both the biological and physical theories were effective and convincing for the purposes for which they were introduced, and both came to dominate their ®elds. Both of them were incomplete. Alternative theories of quantum mechanics with stochastic dynamics are analogous to the mathematical biological theories of the 1940s. In each case the mechanism is clear, but the physical basis of the stochastic ¯uctuations is not. Just as genetic mutation must be slow on the time scale of generations, so the stochastic dynamics of quantum states is slow on the time scales of the SchroÈdinger equation. The processes are much more dif®cult to detect directly than the effects that they produce. As natural selection has led to the development of the beautiful variety of modern species from the simplest beginnings, so the diffusion of quantum states might produce the classical world from the very different quantum world. The quantum theory equivalent of the DNA double helix would be the experimental detection of quantum ¯uctuations on an atomic scale, that is, the detection of the quantum-classical boundary. This will be discussed in the next chapter. 7.4 Quantum theory and physical reality In a letter of 27 May 1926 to SchroÈdinger on his recently proposed wave equation, H. A. Lorentz pointed out that a wave packet, which when moving with the group velocity should represent a particle, `can never stay together and remain con®ned to a small volume in the long run. The slightest dispersion in the medium will pull it apart in the direction of propagation, and even without that dispersion it will always spread more and more in the transverse direction. Because of this unavoidable blurring a wave packet does not seem to be very suitable for representing things that we want to ascribe a rather permanent existence.' This shows the dif®culty of using SchroÈdinger wave packets to represent classical free particles. The dispersion is even stronger when there are particle interactions. This dif®culty is overcome by some of the alternative theories. There are two opposing trends in modern quantum theory. One is pragmatic, and says that it is enough for a theory to answer wellposed questions about experiments and observations. A theory should be

108

Quantum foundations

useful, and no more is required. The other says that we should demand more of our theories, and this was the position of Einstein and of Bell, who wanted a complete description of physical reality [47, 10]. John Bell remained to the end of his life a demanding physicist, who said that quantum mechanics was about beables, and not about observables. He gave support to complete theories, in particular the pilot wave theory of de Broglie and Bohm [18, 84] and the theory of Ghirardi, Rimini and Weber [63, 62]. From Bell's viewpoint these are good quantum theories (GQ) because they are complete and because they satisfy four further conditions [10]: GQ1. The theory can provide a complete and unambiguous description of physical reality. GQ2. The state and evolution of an individual system are explicitly and unambiguously represented. GQ3. There is no ambiguous division between the quantum and classical domains. GQ4. There is no ambiguous division between system and environment.

These conditions are all worth stating, but they are not independent. A theory that satis®es GQ1 must also satisfy the rest, and a theory which satis®es GQ2 must also satisfy GQ3 and GQ4. Ordinary quantum theory does not satisfy these conditions. It does not satisfy GQ3, and so it does not satisfy GQ1 and GQ2. In the quantum theory of measurement, the system evolves according to quantum mechanics, but the apparatus must obey classical laws if it is to provide a record of the measurement. The division between them is ambiguous. Nevertheless, it is our most successful theory. Quantum state diffusion for open systems satis®es GQ3, but not GQ4, because the division between system and environment is ambiguous. As it stands, it is not a good quantum theory in the sense of Bell. Thus QSD provides a different representation of a system for different locations of the boundary between system and environment. That is a practical advantage, because the boundary can be adjusted to represent the behaviour of ensembles of open systems by the simplest theory or computation. The predictions are approximate, as for all theories when they are applied to speci®c problems outside cosmology, because it is not possible to include the whole universe, and anything which doesn't is approximate. But the ambiguity rules out QSD for open systems as a good quantum theory in the sense of Bell. For the example of a quantum measurement, a boundary that is too close to the system does not represent its evolution accurately. A boundary suitable for the computations of chapter 6 gives a good representation of the system, but not necessarily of the measurer. A

7.5 Preparation of quantum states

109

boundary that includes the important `classical' variables of the measurer, but not its microscopic variables, provides a good representation of both quantum system and measurer, because the microscopic variables of the apparatus localize the macroscopic variables. But suppose we take the surface of a sphere s light-years away from the quantum system and measurer as the boundary. There is nothing in principle against this, but the quantum state of the system and the classical variables of the environment then remain entangled until the entanglement can be destroyed by interaction with the environment at least s=2 years later. Until that time the classical variables remain unlocalized, until then there are unphysical `SchroÈdinger cat' states. There is a wide range of choices of boundary between the extremes that gives a reasonable picture of the system and measurer, but the formulation of a fundamental theory should be independent of any such boundary. This is not just a weakness of QSD for open systems, but also of other theories, like those of Zeh [157, 86, 158], and also of OmneÁs form of consistent histories. The decoherent or consistent histories theory [80, 101, 60, 42, 61] satis®es GQ3 but not GQ2. OmneÁs' version does not satisfy GQ4. Nevertheless there are some interesting connections between QSD and the consistent histories approach [40, 81]. If quantum state diffusion is to become a fundamental theory, then the dependence on the boundary must be removed. There must be an intrinsic or spontaneous diffusion of quantum states. It must be weak in systems which SchroÈdinger's equation describes well, and strong for classical systems. It must be universal and independent of any environment. SchroÈdinger's equation must break down somewhere for all physical systems. What contrary evidence do we have that this equation, or its modern extensions to ®elds and strings, applies universally? If the SchroÈdinger equation is to be modi®ed, we need to know how well it has been checked experimentally. How would we check that there are no additional stochastic terms for an arbitrary state of four hydrogen molecules in a 1 cm3 box, with a maximum energy less than the mean energy at room temperature? We could only do so by detecting the coherence properties of the state. The next section shows that this is not just dif®cult, it is impossible, because such an arbitrary state cannot be suf®ciently well prepared. 7.5 Preparation of quantum states A system can never be prepared in an exact quantum state. For example, these days the angle of a polarization plate controls the polarization angle of a photon to no better than about 10ÿ7 radians. A good measure of the

110

Quantum foundations

Figure 7.1 Preparation error cone in Hilbert space.

accuracy of preparation is the Hilbert space angle  between the prepared state j i and the target state j1i: jh j1ij2  cos2 ;

…7:1†

de®ning a cone around j1i in the Hilbert space, as illustrated in ®gure 7.1. The polarization plate is an example of a practical limit to the preparation of a quantum state. There is also a fundamental limit, given by the preparation principle. Preparation principle. If a `classical' system C is used to prepare a state of a quantum system Q, then the number of unambiguous preparation states of Q is no greater than the number of distinguishable quantum states of C. Thus the quantum properties of the `classical' apparatus put a limit on the precision of preparation of a quantum system. The example of the polarization plate makes this clear. If the angular momentum of the plate is suf®ciently well de®ned, then the precision of the angle of orientation is limited by quantum indeterminacy. This is clearly a fundamental limit on the use of the plate for controlling the angle of polarization of photons. The limit becomes much more severe as the number of possible quantum states increases. For general systems, if   =4 the prepared state is useless for testing the interference properties of SchroÈdinger's equation, since the uncertainty in the initial state is too great to test them. Suppose the number N of orthogonal quantum states of the system is large. Then the numerical factors in the

7.5 Preparation of quantum states

111

following can be neglected. Use a logarithmic measure of the number of cells of a partition. This is an entropy in dimensionless units. Measurement

Preparation Hilbert entropy SH

Boltzmann entropy SB Partition phase space into cells (Boltzmann: arbitrary size) (Modern: Planck cells) Number of cells = N (distinct measurement states)

Partition sphere of normed states into cells Number of cells  ÿ2N (distinct preparation states)

Putting these relations together we ®nd that the Hilbert entropy for the preparation of a system is approximately equal to the exponential of its Boltzmann entropy: SB ˆ ln N;

SH  ÿ2N ln   2j ln jeSB :

…7:2†

We now use the preparation principle, which can be expressed in terms of entropies of the quantum system Q and the classical system C that is used to prepare it as ) SH …Q†  SB …C†; …7:3† (preparation principle). SB …Q†  ln SB …C† The observable universe is an upper bound on the size of the system C used to prepare Q. Its Boltzmann entropy is SB …universe†  1088 , so if this is used to prepare a quantum system Q, then SB …Q†  88 ln 10  200:

…7:4†

Figure 7.2 illustrates four hydrogen molecules in a box. The wave number of an H2 molecule with the mean kinetic energy for a temperature of 300 K is about 3:5  1010 mÿ1 , giving a quantum number of about n ˆ 108 for translational motion within a centimetre in one direction. This is illustrated in ®gure 7.2. So for all three directions there are roughly n3  1024 states below this energy, and a resultant Boltzmann entropy of translational motion SB ˆ 55 in the dimensionless units used in this section. Consequently four molecules have a greater Boltzmann entropy than 200, and it is impossible to prepare an

112

Quantum foundations

Figure 7.2 States of H2 molecules in a 1 cm3 box.

arbitrary state of this system suf®ciently well to test its interference properties [110, 129]. Suppose that quantum state diffusion were an intrinsic property of isolated quantum systems, and that it was signi®cant for most quantum states of these hydrogen molecules. For most states we have no way to detect the effect of the diffusion on the molecules. We could detect it for some unentangled systems, but there are relatively few of these, and it could be absent for them. The experimental evidence for an unmodi®ed SchroÈdinger equation is very weak, even for small systems like a few atoms, let alone systems that are normally considered to be classical. There is plenty of room to modify it. The problem is to detect whatever modi®cation there might be. 7.6 Too many alternative theories Here is a brief account of the early development of the alternative theories that are most closely related to quantum state diffusion. It is not an account of all alternative theories, which would occupy more than the whole book. It

7.6 Too many alternative theories

113

does not include theories based on gravity or on spacetime ¯uctuations, which appear in the next chapter. In 1966 Bohm and Bub [17] proposed a theory with a space of ordinary state ket vectors j i, which represent the state of quantum systems, and satisfy SchroÈdinger's equation. There is also a dual space of bra vectors which is different from the state vector space and has different dynamics. The components of the bra vectors are to introduce the stochasticity into the physics, but the details of the dynamics are unspeci®ed. Bohm and Bub pointed out that measuring apparatus was just another kind of environment for a quantum system, and should be treated like other environments. In the same year Nelson put forward his `stochastic quantum mechanics' in which quantum indeterminacy comes from a sort of Brownian motion of particles in position space on the scale of the usual wave packets [98, 99]. He used the Itoà calculus. In 1976 Pearle [103] introduced nonlinear terms into the SchroÈdinger equation for an individual quantum system, in which apparently random variations are produced by small changes in the phase factor of the state. In a measurement there is a resulting diffusion towards the eigenstates. In 1979 [104] he wrote an article with the title `Towards explaining why events occur', which very clearly stated the need for reality of the waves of quantum dynamics, and introduced an explicit stochastic equation for these waves. In 1984, Gisin [68, 67] considered the measurement of a two-state system, represented by the projector P, and contrasted the von Neumann postulate for the density operator  ! PP ‡ …I ÿ P†…I ÿ P†;

…7:5†

which is a linear relation, with the LuÈders postulate for the state vector ) j i ! jP i=jjP jj with probability h jPj i …7:6† j i ! j…I ÿ P† i=jj…I ÿ P† jj with probability h jI ÿ Pj i; which is nonlinear. He showed that LuÈders transition follows from the long time limit of the solution of a nonlinear stochastic Itoà equation of QSD form, but without the complex ¯uctuations, demonstrating localization to an eigenstate. He derived the von Neumann postulate as a limit of the linear master equation for the density operator obtained from the ¯uctuating pure states. These letters demonstrated the importance of a consistent evolution of the master equation, which he analysed in more detail in 1989 [69], as described

114

Quantum foundations

in section 7.7. In 1986 DioÂsi [32] showed how to unravel a dissipative master equation using quantum jumps. Also in 1986, Ghirardi, Rimini and Weber [63] produced the ®rst complete theory, with a universal stochastic localization of position. They demonstrated how the localization could be undetectable for small quantum systems, yet dominant for large classical ones. The rates were speci®ed by two undetermined parameters. However, the theory was based on equations for density operators, not pure states, and on jumps. Bell [10] presented the theory in terms of stochastic equations for the states, and it was later reformulated in terms of diffusing quantum states by DioÂsi [34], Gisin [69], Pearle [107] and Ghirardi, Pearle and Rimini [64], who gave it the name of continuous spontaneous localization. In 1988 DioÂsi extended Gisin's state diffusion theory to include position localization. He showed that a free particle localizes to a Gaussian wave packet that behaves like a Brownian particle [34, 35]. He also generalized the theory for wide open systems in terms of the probability distribution Pr… † of section 3.2, with an alternative representation in terms of orthogonal quantum jumps [36]. These theories all suppose some intrinsic, spontaneous diffusion or jumps, and satisfy Bell's conditions. They all have real ¯uctuations in the sense of section 4.6, not the complex unitary invariant ¯uctuations of QSD introduced in [72] and used in chapter 4 to obtain the QSD equation from the master equation. The theories which follow the early work of Ghirardi, Rimini and Weber are based on Lindblads which depend on the position. However, there is strong numerical and other evidence from QSD that phase space localization to wave packets is a property of all systems with both a Hamiltonian and a suf®ciently strong Hermitian Lindblad or Lindblads, except for very simple or special cases. The Lindblad produces localization in only one direction in the 2d-dimensional phase space, but the Hamiltonian then mixes this direction with the others, so that the quantum wave localizes in all directions in the phase space. This process is very clearly demonstrated for the example in section 4.3 of a mechanical oscillator whose Lindblad is proportional to the momentum. It localizes in both momentum and position, producing an oscillating wave packet with near minimum indeterminacy product. A set of mechanical oscillators like atoms in a solid, with momentum localization, will localize in phase space, and therefore in position. Collisions between particles have the same effect, so atoms in gases with a variety of Hermitian Lindblads will also localize in phase space. Phase space localization includes position localization, so, for real systems with Hamiltonians

7.7 Gisin condition

115

that include interaction, position localization does not need Lindblads that depend on position. The energy localization of the next chapter is very special. For nonrelativistic ¯uctuations, it does not localize in position. Intrinsic state diffusion could occur for entangled systems as simple as four hydrogen molecules at 300 K without violating any possible experiments, let alone actual experiments. Conseqently there is an enormous range of Lindblads and parameters for intrinsic QSD that are consistent with experience and satisfy Bell's conditions. The problem is not to ®nd an alternative theory, but to choose among this variety of possible theories. How then can we guide experimenters to possible violations of SchroÈdinger's equation so that they can detect the quantumclassical boundary? The next section describes an important constraint on possible theories, but the variety remains. The next chapter suggests an answer. 7.7 Gisin condition Shimony [134] uses `peaceful coexistence' to describe the resolution of the dif®cult problems of compatibility of the nonlocality of quantum mechanics and special relativity. In the next chapter we show that not all of them have been resolved. A condition on compatibility of realistic nonlinear evolution equations and special relativity was put forward by Gisin [68] as part of a lively exchange of comments with Pearle [105], who subsequently gave a detailed treatment of the special case of matrix diagonalization [106]. This section mainly follows Gisin's 1989 paper [69], whose methods he [70] and Polchinski [122] used in a critique of Weinberg's nonlinear quantum theory [151, 152]. In chapters 3 and 4 we accepted the experimental evidence that the evolution of an ensemble of quantum systems depends on the density operator q and not on its unravelling into pure states. We now show that this fundamental property is a consequence of special relativity, that signals cannot be transmitted faster than the velocity of light, and also an additional assumption about preparation of entangled systems. It follows from these conditions, and from the linearity of the evolution of density operators, that any deterministic evolution of the state vectors must also be linear. Suppose the density operator q…0† of a system A at time t ˆ 0 has two unravellings into pure states jii with projectors Pi …0† and into states j ji with projectors Pj …0†, where there are m different values of i and n  m different values of j

116

q…0† ˆ

X i

Quantum foundations

Pr…i†Pi …0† ˆ

X j

Pr…j†Pj …0†

…Pr…i† 6ˆ 0; Pr…j† 6ˆ 0†:

…7:7†

The states with projectors Pi …0† need not be orthogonal, and the same is true of the Pj …0†. None of the probabilities Pr…i†; Pr…j† is zero. Since the sets of states jii and j ji make the same density operator they must span the same state space, and since none of the probabilities is zero, we can ®nd real amplitudes ci ; cj and a matrix Vij such that )

c2i ˆ Pr…i†; c2j ˆ Pr…j†; X Vij cj j ji: ci jii ˆ

…7:8†

j

Substituting into the equations (7.7) for the density operator then gives X ijj 0

cj Vij j jih j 0 jVj0 i cj 0 ˆ

X j

c2j j jih jj;

so

X i

Vj0 i Vij ˆ j 0 j : …7:9†

For m ˆ n the matrix Vij is unitary, and in general it is a rectangular generalization of a unitary matrix. Suppose that there is a distant physical system B with a set of m orthonormal states jui i and another set of n states jvj i ˆ

X i

Vji jui i;

…7:10†

whose orthonormality follows from the generalized unitary property (7.9) of Vij . Then the entangled state ji ˆ

X i

ci jiijui i ˆ

X j

cj j jijvj i

…7:11†

of the combined system AB has a reduced density operator qA …0† ˆ TrB jihj:

…7:12†

If jui i and jvj i are nondegenerate eigenstates of two operators Y and N for the system B, then a Y measurement of B unravels the system A in one way and an N measurement unravels it the other way. Now suppose, contrary to what we want to prove, that the evolution of pure states is deterministic and nonlinear, and that the evolution of the two

7.7 Gisin condition

117

unravellings can lead to different density operators at a later time t. Denote the deterministic evolution by the possibly nonlinear (super)operator T…t† on the projectors P such that T…t†P…0† ˆ P…t†:

…7:13†

Suppose that the two ensembles corresponding to the two unravellings evolve after time t into ensembles with different density operators q1 …t† ˆ

X i

T…t†Pi …0† 6ˆ q2 …t† ˆ

X j

T…t†Pj …0†:

…7:14†

Suppose the entangled system ji can be prepared with A and B further apart than ct where c is the velocity of light. Suppose an experimenter Barbara at B wants to transmit a binary message to an experimenter Arthur at A faster than the velocity of light: `yes, I will', or `no, I won't'. Then she has to measure Y for yes or N for no at time 0. The result is that the system A goes into one or other of the unravelled ensembles. Arthur patiently waits for a time t, and then measures the expectation of an operator G for which TrGq1 …t† 6ˆ TrGq2 …t†:

…7:15†

The probability of a given g depends on the unravelling, and so it depends on Barbara's message. This changes the probability of `yes' and of `no', but doesn't give a de®nite answer. The probability of an unfortunate misunderstanding can be reduced to an arbitrarily small value by carrying out many copies of the experiment in parallel. Physics prevents even the most passionate messages being sent faster than the velocity of light, and so if the state  can be prepared, it follows that the evolution of the density operators cannot depend on the unravelling. As a direct result of this it follows that the operator T…t† of (7.13) applies also to the density operators, whose evolution is linear. Consequently the evolution of the projectors must be linear, and that means that the evolution of the states is also linear. We conclude with Gisin that if signals cannot be sent faster than the velocity of light, and if entangled states of the type ji can be prepared, then deterministic nonlinear evolution of state vectors is impossible. We have seen in section 7.5 that there are strong limitations on the preparation of entangled states for all but the simplest systems, so we can only

118

Quantum foundations

conclude that the deterministic nonlinear evolution contradicts special relativity for these. But it would be dif®cult to construct a theory with strictly linear evolution for simple systems and nonlinear evolution for more complicated systems, so, whilst deterministic nonlinear evolution of state vectors is not entirely ruled out, it is unlikely. Gisin and Rigo have taken this argument further, by showing how to add precise noise terms to the deterministic nonlinear evolution equation which give solutions closest to the solutions of the master equation [76].

119

8 Primary state diffusion ± PSD 8.1 8.2 8.3 8.4 8.5 8.6 8.7

First approach ± SchroÈdinger from diffusion Decoherence Feynman's lectures on gravitation Second approach ± spacetime PSD Geometry of the ¯uctuations Matter interferometers Conclusions

Primary state diffusion, or PSD, is obtained twice, from different principles. First comes a nonrelativistic formulation that follows from the need to derive the long-time behaviour of a physical system from its short-time behaviour [112, 113]. The second formulation relates PSD to spacetime ¯uctuations, and follows a long tradition that relates alternative quantum theories to gravity [114, 115]. The second formulation has no free parameters, but it is not unique, because of different possible geometries of ¯uctuations in spacetime. Experimental tests are proposed, which might also serve as tests of quantum gravity theories. The chapter concludes with the implications for quantum measurement and a discussion of the many remaining unsolved problems. Some of these issues are discussed in [116, 115]. 8.1 First approach ± SchroÈdinger from diffusion The ®rst form of primary state diffusion is derived in this section as the simplest formulation of a nonrelativistic theory in which quantum state diffusion is a fundamental property of all physical systems [112]. This theory leads to energy localization, a good approximation for systems that remain suf®ciently localized in space, but the nonrelativistic theory is incomplete because there is no space localization of extended systems. In science, the evolution of a system over a long time period can be obtained by combining the effects over shorter time periods. If diffusion and drift are both present, then for short periods the diffusion dominates the drift. If they are suf®ciently short, the drift is negligible compared with the diffusion. This leads to a reversal of the usual roles of diffusion and SchroÈdinger dynamics, whereby the long-time SchroÈdinger dynamics is derived from the short-term diffusion dynamics. The diffusion is fundamen-

120

Primary state diffusion ± PSD

tal and universal, and we derive the drift from the diffusion, not the other way round. Primary state diffusion means that the quantum state diffusion is primary and the drift, including ordinary quantum mechanics, is secondary. In the ®rst version, the principles of primary state diffusion (PSD) are: PSD1. State vector. The state of a system at any time t is represented by a normalized state vector j …t†i. This applies to classical systems like the solar system, pointers, data records and Brownian particles as well as to quantum systems like electrons and photons. PSD2. Operator. The dynamics of the system is determined by a system diffusion operator K, which is linear. PSD3. State diffusion. The change of state obeys an elementary QSD equation, with Lindblad K. There is no SchroÈdinger term.

jd i ˆ hKi K ÿ 12Ky K ÿ 12hKi hKi†j idt ‡ K
j id:

…8:1†

The operator K is not Hermitian and K
ˆ K ÿ hKi, as in equation (3.22). The quantum ¯uctuations of PSD3 are independent of spatial coordinates in the nonrelativistic theory, but are space-dependent in any relativistic theory. The operator K determines the SchroÈdinger dynamics along with the state diffusion. The choice of K is severely constrained by the following conditions. C1. SchroÈdinger evolution. Over longer periods, when state diffusion is negligible, the SchroÈdinger evolution is given by the anti-Hermitian part of the drift operator in the QSD equation, so the quantum state diffusion equation must be of the form

jd i ˆ ÿ…i= hÿ†Hj idt ‡ Rj idt ‡ K
j id;

…8:2†

where the ®rst term represents the regular SchroÈdinger evolution and the second term is the residual drift. C2. Limited disruption. The diffusion is not so fast that it disrupts the regular SchroÈdinger evolution to a degree that violates experimental or observational evidence. C3. Fast diffusion. The rate of diffusion is suf®ciently fast to localize classical dynamical variables.

8.1 First approach ± SchroÈdinger from diffusion

121

The last condition is based on the assumption that there are dynamical variables that behave classically, and that this classical behaviour does not normally depend on acts of observation or measurement. For systems described by classical variables, the state vectors must be so localized in con®guration and phase space that they cannot be distinguished from classical points which move according to the classical equations of motion. In particular the motion of the centre of mass of a planet, stone or Brownian particle is classical. Notice that conditions C2 and C3 are opposing restrictions on PSD theory, one ensuring that the state diffusion is not too fast, and the other that it is not too slow. By equating corresponding terms in the state diffusion equation of PSD3 and equation (8.2) of C1, we ®nd that for all j i, h jKR j iKI ÿ h jKI j iKR ˆ c… †I ÿ H=hÿ ˆ c… †I ÿ h jIj iH=hÿ;

…8:3†

R… † ˆ ÿ12 Ky
… †K
… †;

…8:4†

where I is the unit operator, c… † is a real constant corresponding to an arbitrary phase factor, and KR ; KI are Hermitian. So c… † must have the form c… † ˆ h jCj i

…8:5†

for some operator C. Since the equation is valid for all j i, it can be written as an equation for outer products of operators, as C I ÿ I H= hÿ ˆ KR KI ÿ KI KR :

…8:6†

Since the right side is antisymmetric, so is the left side, and …H= hÿ† I ÿ I …H= hÿ† ˆ KR KI ÿ KI KR :

…8:7†

K must be a linear combination of the Hamiltonian and the unit operator I, and since the dynamics is independent of a phase factor of unit modulus multiplying K, the coef®cient a1 of H can be taken as positive. Substitution shows that the real part of the coef®cient of I has no effect on the dynamics, so

122

Primary state diffusion ± PSD

K ˆ a1 H ‡ ia2 I;

…8:8†

where a1 is positive and a2 is real, so that by (8.7) i K ˆ a1 H ‡ ÿ I: ha1

…8:9†

When this is substituted into the original QSD equation it leads to a transformed QSD equation jd i ˆ …ÿi= hÿ†H
dt ÿ a21 H2
j idt ‡ a1 H
j id

…transformed†:

…8:10†

Note that as a1 ! 0, the diffusion becomes negligible and the transformed QSD is the SchroÈdinger equation, apart from a nonphysical phase factor. This equation of nonrelativistic PSD has the same form as the state diffusion equation for the measurement of energy. So the principles PSD1±PSD3 and the condition C1 lead to a universal intrinsic energy localization. This was ®rst proposed by Bedford and Wang [7, 8], tentatively suggested as an intrinsic state diffusion process by Gisin [69], and appeared as an approximation to a density operator decoherence theory in the work of Milburn [95] and Moya-Cessa et al. [97]. In the extreme case of a single Hamiltonian for the whole universe, the ¯uctuation d acts simultaneously over all space at a given time. This satis®es Galilean relativity, but is an obvious violation of special relativity, for which there is no absolute simultaneity. Nevertheless the nonrelativistic theory gives a good picture of what happens to a system of nonzero rest mass in a limited region of spacetime. In PSD, as in Brownian motion, there is a time constant 0 that marks the boundary between the time intervals for which diffusion dominates and those for which the drift dominates. This is given by 0 ˆ a21 hÿ2 ;

a1 ˆ 01=2 = hÿ:

The time constant 0 is universal for PSD, and more useful than a1 .

…8:11†

8.2 Decoherence

123

8.2 Decoherence Ramsey devised a method of putting an atom into a coherent linear combination of two different energy states with energy difference
E ˆ hÿ!. Since PSD produces energy localization, these combinations have a fundamental characteristic lifetime, after which the diffusion reduces the standard deviation in energy to less than
E, so that the system tends to go into one or other of the energy states. Using the general theory of localization, this gives a `primary decoherence rate' ÿP , ÿP ˆ !2 0 ;

…8:12†

whose inverse is a primary coherence decay time. The decay time increases as the square of the energy difference. If a Ramsey experiment does not show decoherence, this puts an upper bound on 0 . This leads to a theory of decoherence rates. Primary state diffusion is not the only possible reason for decoherence. It can be dissipated earlier by decay, like the radiative decay of an atom, or the radioactive decay of a nucleus, or by interaction with matter in the environment, as in line broadening, after which the coherence is no longer observable. Normal environments are suf®ciently big for this secondary decoherence by dissipation to mask the primary decoherence, and the rate for this will be denoted ÿS , the secondary coherence decay rate. Experiments are not normally designed to measure decoherence rates, but nevertheless many do so, because unwanted decoherence interferes with the purpose of the experiments. Because all relevant experiments to date have been found consistent with ordinary quantum theory, ÿS > ÿP

…8:13†

and 0